| blob | 296d414c812a261a2d1a59b56f67d4de53039685 |
1 /*
2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
3 *
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 *
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
16 * distribution.
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.60 2008/07/01 16:48:51 dillon Exp $
35 */
37 /*
38 * HAMMER B-Tree index
39 *
40 * HAMMER implements a modified B+Tree. In documentation this will
41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
42 * looks like a B+Tree (A B-Tree which stores its records only at the leafs
43 * of the tree), but adds two additional boundary elements which describe
44 * the left-most and right-most element a node is able to represent. In
45 * otherwords, we have boundary elements at the two ends of a B-Tree node
46 * instead of sub-tree pointers.
47 *
48 * A B-Tree internal node looks like this:
49 *
50 * B N N N N N N B <-- boundary and internal elements
51 * S S S S S S S <-- subtree pointers
52 *
53 * A B-Tree leaf node basically looks like this:
54 *
55 * L L L L L L L L <-- leaf elemenets
56 *
57 * The radix for an internal node is 1 less then a leaf but we get a
58 * number of significant benefits for our troubles.
59 *
60 * The big benefit to using a B-Tree containing boundary information
61 * is that it is possible to cache pointers into the middle of the tree
62 * and not have to start searches, insertions, OR deletions at the root
63 * node. In particular, searches are able to progress in a definitive
64 * direction from any point in the tree without revisting nodes. This
65 * greatly improves the efficiency of many operations, most especially
66 * record appends.
67 *
68 * B-Trees also make the stacking of trees fairly straightforward.
69 *
70 * INSERTIONS: A search performed with the intention of doing
71 * an insert will guarantee that the terminal leaf node is not full by
72 * splitting full nodes. Splits occur top-down during the dive down the
73 * B-Tree.
74 *
75 * DELETIONS: A deletion makes no attempt to proactively balance the
76 * tree and will recursively remove nodes that become empty. If a
77 * deadlock occurs a deletion may not be able to remove an empty leaf.
78 * Deletions never allow internal nodes to become empty (that would blow
79 * up the boundaries).
80 */
82 #include <sys/buf.h>
83 #include <sys/buf2.h>
93 /*
94 * Iterate records after a search. The cursor is iterated forwards past
95 * the current record until a record matching the key-range requirements
96 * is found. ENOENT is returned if the iteration goes past the ending
97 * key.
98 *
99 * The iteration is inclusive of key_beg and can be inclusive or exclusive
100 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
101 *
102 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
103 * may be modified by B-Tree functions.
104 *
105 * cursor->key_beg may or may not be modified by this function during
106 * the iteration. XXX future - in case of an inverted lock we may have
107 * to reinitiate the lookup and set key_beg to properly pick up where we
108 * left off.
109 *
110 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
111 */
112 int
114 {
115 hammer_node_ondisk_t node;
116 hammer_btree_elm_t elm;
121 /*
122 * Skip past the current record
123 */
130 }
132 /*
133 * Loop until an element is found or we are done.
134 */
136 /*
137 * We iterate up the tree and then index over one element
138 * while we are at the last element in the current node.
139 *
140 * If we are at the root of the filesystem, cursor_up
141 * returns ENOENT.
142 *
143 * XXX this could be optimized by storing the information in
144 * the parent reference.
145 *
146 * XXX we can lose the node lock temporarily, this could mess
147 * up our scan.
148 */
159 curthread);
160 }
161 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
165 /* reload stale pointer */
169 /*
170 * If we are reblocking we want to return internal
171 * nodes.
172 */
176 }
179 }
181 /*
182 * Check internal or leaf element. Determine if the record
183 * at the cursor has gone beyond the end of our range.
184 *
185 * We recurse down through internal nodes.
186 */
200 r,
201 curthread
202 );
210 s
211 );
212 }
217 }
222 }
225 /*
226 * Better not be zero
227 */
230 /*
231 * If running the mirror filter see if we can skip
232 * the entire sub-tree.
233 */
239 }
240 }
246 /* reload stale pointer */
262 r
263 );
264 }
268 }
270 /*
271 * We support both end-inclusive and
272 * end-exclusive searches.
273 */
278 }
286 }
291 }
294 }
295 /*
296 * node pointer invalid after loop
297 */
299 /*
300 * Return entry
301 */
311 );
312 }
314 }
316 }
318 /*
319 * Iterate in the reverse direction. This is used by the pruning code to
320 * avoid overlapping records.
321 */
322 int
324 {
325 hammer_node_ondisk_t node;
326 hammer_btree_elm_t elm;
331 /*
332 * Skip past the current record. For various reasons the cursor
333 * may end up set to -1 or set to point at the end of the current
334 * node. These cases must be addressed.
335 */
342 }
346 /*
347 * Loop until an element is found or we are done.
348 */
353 /*
354 * We iterate up the tree and then index over one element
355 * while we are at the last element in the current node.
356 */
362 }
363 /* reload stale pointer */
368 }
370 /*
371 * Check internal or leaf element. Determine if the record
372 * at the cursor has gone beyond the end of our range.
373 *
374 * We recurse down through internal nodes.
375 */
389 r
390 );
398 s
399 );
400 }
405 }
408 /*
409 * Better not be zero
410 */
417 /* reload stale pointer */
420 /* this can assign -1 if the leaf was empty */
436 s
437 );
438 }
442 }
450 }
455 }
458 }
459 /*
460 * node pointer invalid after loop
461 */
463 /*
464 * Return entry
465 */
475 );
476 }
478 }
480 }
482 /*
483 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
484 * could not be found, EDEADLK if inserting and a retry is needed, and a
485 * fatal error otherwise. When retrying, the caller must terminate the
486 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
487 *
488 * The cursor is suitably positioned for a deletion on success, and suitably
489 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
490 * specified.
491 *
492 * The cursor may begin anywhere, the search will traverse the tree in
493 * either direction to locate the requested element.
494 *
495 * Most of the logic implementing historical searches is handled here. We
496 * do an initial lookup with create_tid set to the asof TID. Due to the
497 * way records are laid out, a backwards iteration may be required if
498 * ENOENT is returned to locate the historical record. Here's the
499 * problem:
500 *
501 * create_tid: 10 15 20
502 * LEAF1 LEAF2
503 * records: (11) (18)
504 *
505 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
506 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
507 * not visible and thus causes ENOENT to be returned. We really need
508 * to check record 11 in LEAF1. If it also fails then the search fails
509 * (e.g. it might represent the range 11-16 and thus still not match our
510 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
511 * further iterations.
512 *
513 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
514 * and the cursor->create_check TID if an iteration might be needed.
515 * In the above example create_check would be set to 14.
516 */
517 int
519 {
531 /*
532 * Stop if no error.
533 * Stop if error other then ENOENT.
534 * Stop if ENOENT and not special case.
535 */
537 }
541 }
543 /* loop */
544 }
547 }
551 }
553 /*
554 * Execute the logic required to start an iteration. The first record
555 * located within the specified range is returned and iteration control
556 * flags are adjusted for successive hammer_btree_iterate() calls.
557 */
558 int
560 {
567 }
570 }
572 /*
573 * Similarly but for an iteration in the reverse direction.
574 *
575 * Set ATEDISK when iterating backwards to skip the current entry,
576 * which after an ENOENT lookup will be pointing beyond our end point.
577 */
578 int
580 {
592 }
595 }
597 /*
598 * Extract the record and/or data associated with the cursor's current
599 * position. Any prior record or data stored in the cursor is replaced.
600 * The cursor must be positioned at a leaf node.
601 *
602 * NOTE: All extractions occur at the leaf of the B-Tree.
603 */
604 int
606 {
607 hammer_mount_t hmp;
608 hammer_node_ondisk_t node;
609 hammer_btree_elm_t elm;
610 hammer_off_t data_off;
614 /*
615 * The case where the data reference resolves to the same buffer
616 * as the record reference must be handled.
617 */
623 /*
624 * There is nothing to extract for an internal element.
625 */
629 /*
630 * Only record types have data.
631 */
644 /*
645 * Load the data
646 */
653 }
656 /*
657 * Insert a leaf element into the B-Tree at the current cursor position.
658 * The cursor is positioned such that the element at and beyond the cursor
659 * are shifted to make room for the new record.
660 *
661 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
662 * flag set and that call must return ENOENT before this function can be
663 * called.
664 *
665 * The caller may depend on the cursor's exclusive lock after return to
666 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
667 *
668 * ENOSPC is returned if there is no room to insert a new record.
669 */
670 int
672 {
673 hammer_node_ondisk_t node;
681 /*
682 * Insert the element at the leaf node and update the count in the
683 * parent. It is possible for parent to be NULL, indicating that
684 * the filesystem's ROOT B-Tree node is a leaf itself, which is
685 * possible. The root inode can never be deleted so the leaf should
686 * never be empty.
687 *
688 * Remember that the right-hand boundary is not included in the
689 * count.
690 */
700 }
704 /*
705 * Update the leaf node's aggregate mirror_tid for mirroring
706 * support.
707 */
714 /*
715 * What we really want to do is propogate mirror_tid all the way
716 * up the parent chain to the B-Tree root. That would be
717 * ultra-expensive, though.
718 */
724 }
726 /*
727 * Debugging sanity checks.
728 */
733 }
738 }
740 /*
741 * Delete a record from the B-Tree at the current cursor position.
742 * The cursor is positioned such that the current element is the one
743 * to be deleted.
744 *
745 * On return the cursor will be positioned after the deleted element and
746 * MAY point to an internal node. It will be suitable for the continuation
747 * of an iteration but not for an insertion or deletion.
748 *
749 * Deletions will attempt to partially rebalance the B-Tree in an upward
750 * direction, but will terminate rather then deadlock. Empty internal nodes
751 * are never allowed by a deletion which deadlocks may end up giving us an
752 * empty leaf. The pruner will clean up and rebalance the tree.
753 *
754 * This function can return EDEADLK, requiring the caller to retry the
755 * operation after clearing the deadlock.
756 */
757 int
759 {
760 hammer_node_ondisk_t ondisk;
761 hammer_node_t node;
762 hammer_node_t parent;
770 /*
771 * Delete the element from the leaf node.
772 *
773 * Remember that leaf nodes do not have boundaries.
774 */
785 }
789 /*
790 * Validate local parent
791 */
797 }
799 /*
800 * If the leaf becomes empty it must be detached from the parent,
801 * potentially recursing through to the filesystem root.
802 *
803 * This may reposition the cursor at one of the parent's of the
804 * current node.
805 *
806 * Ignore deadlock errors, that simply means that btree_remove
807 * was unable to recurse and had to leave us with an empty leaf.
808 */
816 }
820 }
822 /*
823 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
824 *
825 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
826 *
827 * The search can begin ANYWHERE in the B-Tree. As a first step the search
828 * iterates up the tree as necessary to properly position itself prior to
829 * actually doing the sarch.
830 *
831 * INSERTIONS: The search will split full nodes and leaves on its way down
832 * and guarentee that the leaf it ends up on is not full. If we run out
833 * of space the search continues to the leaf (to position the cursor for
834 * the spike), but ENOSPC is returned.
835 *
836 * The search is only guarenteed to end up on a leaf if an error code of 0
837 * is returned, or if inserting and an error code of ENOENT is returned.
838 * Otherwise it can stop at an internal node. On success a search returns
839 * a leaf node.
840 *
841 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
842 * filesystem, and it is not simple code. Please note the following facts:
843 *
844 * - Internal node recursions have a boundary on the left AND right. The
845 * right boundary is non-inclusive. The create_tid is a generic part
846 * of the key for internal nodes.
847 *
848 * - Leaf nodes contain terminal elements only now.
849 *
850 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
851 * historical search. ASOF and INSERT are mutually exclusive. When
852 * doing an as-of lookup btree_search() checks for a right-edge boundary
853 * case. If while recursing down the left-edge differs from the key
854 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
855 * with cursor->create_check. This is used by btree_lookup() to iterate.
856 * The iteration backwards because as-of searches can wind up going
857 * down the wrong branch of the B-Tree.
858 */
859 static
860 int
862 {
863 hammer_node_ondisk_t node;
864 hammer_btree_elm_t elm;
883 curthread
884 );
896 );
897 }
899 /*
900 * Move our cursor up the tree until we find a node whos range covers
901 * the key we are trying to locate.
902 *
903 * The left bound is inclusive, the right bound is non-inclusive.
904 * It is ok to cursor up too far.
905 */
916 }
918 /*
919 * The delete-checks below are based on node, not parent. Set the
920 * initial delete-check based on the parent.
921 */
926 }
928 /*
929 * We better have ended up with a node somewhere.
930 */
933 /*
934 * If we are inserting we can't start at a full node if the parent
935 * is also full (because there is no way to split the node),
936 * continue running up the tree until the requirement is satisfied
937 * or we hit the root of the filesystem.
938 *
939 * (If inserting we aren't doing an as-of search so we don't have
940 * to worry about create_check).
941 */
949 }
953 }
956 /* node may have become stale */
959 }
961 /*
962 * Push down through internal nodes to locate the requested key.
963 */
966 /*
967 * Scan the node to find the subtree index to push down into.
968 * We go one-past, then back-up.
969 *
970 * We must proactively remove deleted elements which may
971 * have been left over from a deadlocked btree_remove().
972 *
973 * The left and right boundaries are included in the loop
974 * in order to detect edge cases.
975 *
976 * If the separator only differs by create_tid (r == 1)
977 * and we are doing an as-of search, we may end up going
978 * down a branch to the left of the one containing the
979 * desired key. This requires numerous special cases.
980 */
986 }
988 /*
989 * Try to shortcut the search before dropping into the
990 * linear loop. Locate the first node where r <= 1.
991 */
1000 }
1007 }
1009 }
1013 }
1015 /*
1016 * These cases occur when the parent's idea of the boundary
1017 * is wider then the child's idea of the boundary, and
1018 * require special handling. If not inserting we can
1019 * terminate the search early for these cases but the
1020 * child's boundaries cannot be unconditionally modified.
1021 */
1023 /*
1024 * If i == 0 the search terminated to the LEFT of the
1025 * left_boundary but to the RIGHT of the parent's left
1026 * boundary.
1027 */
1028 u_int8_t save;
1032 /*
1033 * If we aren't inserting we can stop here.
1034 */
1039 }
1041 /*
1042 * Correct a left-hand boundary mismatch.
1043 *
1044 * We can only do this if we can upgrade the lock,
1045 * and synchronized as a background cursor (i.e.
1046 * inserting or pruning).
1047 *
1048 * WARNING: We can only do this if inserting, i.e.
1049 * we are running on the backend.
1050 */
1061 /*
1062 * If i == node->count + 1 the search terminated to
1063 * the RIGHT of the right boundary but to the LEFT
1064 * of the parent's right boundary. If we aren't
1065 * inserting we can stop here.
1066 *
1067 * Note that the last element in this case is
1068 * elms[i-2] prior to adjustments to 'i'.
1069 */
1075 }
1077 /*
1078 * Correct a right-hand boundary mismatch.
1079 * (actual push-down record is i-2 prior to
1080 * adjustments to i).
1081 *
1082 * We can only do this if we can upgrade the lock,
1083 * and synchronized as a background cursor (i.e.
1084 * inserting or pruning).
1085 *
1086 * WARNING: We can only do this if inserting, i.e.
1087 * we are running on the backend.
1088 */
1099 /*
1100 * The push-down index is now i - 1. If we had
1101 * terminated on the right boundary this will point
1102 * us at the last element.
1103 */
1105 }
1113 i,
1119 );
1120 }
1122 /*
1123 * We better have a valid subtree offset.
1124 */
1127 /*
1128 * Handle insertion and deletion requirements.
1129 *
1130 * If inserting split full nodes. The split code will
1131 * adjust cursor->node and cursor->index if the current
1132 * index winds up in the new node.
1133 *
1134 * If inserting and a left or right edge case was detected,
1135 * we cannot correct the left or right boundary and must
1136 * prepend and append an empty leaf node in order to make
1137 * the boundary correction.
1138 *
1139 * If we run out of space we set enospc and continue on
1140 * to a leaf to provide the spike code with a good point
1141 * of entry.
1142 */
1150 }
1151 /*
1152 * reload stale pointers
1153 */
1156 }
1157 }
1159 /*
1160 * Push down (push into new node, existing node becomes
1161 * the parent) and continue the search.
1162 */
1164 /* node may have become stale */
1168 }
1170 /*
1171 * We are at a leaf, do a linear search of the key array.
1172 *
1173 * On success the index is set to the matching element and 0
1174 * is returned.
1175 *
1176 * On failure the index is set to the insertion point and ENOENT
1177 * is returned.
1178 *
1179 * Boundaries are not stored in leaf nodes, so the index can wind
1180 * up to the left of element 0 (index == 0) or past the end of
1181 * the array (index == node->count). It is also possible that the
1182 * leaf might be empty.
1183 */
1191 }
1193 /*
1194 * Try to shortcut the search before dropping into the
1195 * linear loop. Locate the first node where r <= 1.
1196 */
1207 /*
1208 * We are at a record element. Stop if we've flipped past
1209 * key_beg, not counting the create_tid test. Allow the
1210 * r == 1 case (key_beg > element but differs only by its
1211 * create_tid) to fall through to the AS-OF check.
1212 */
1220 }
1222 /*
1223 * Check our as-of timestamp against the element.
1224 */
1230 }
1231 /* success */
1236 }
1237 /* success */
1238 }
1244 }
1246 }
1248 /*
1249 * The search of the leaf node failed. i is the insertion point.
1250 */
1251 failed:
1255 }
1257 /*
1258 * No exact match was found, i is now at the insertion point.
1259 *
1260 * If inserting split a full leaf before returning. This
1261 * may have the side effect of adjusting cursor->node and
1262 * cursor->index.
1263 */
1272 }
1273 /*
1274 * reload stale pointers
1275 */
1276 /* NOT USED
1277 i = cursor->index;
1278 node = &cursor->node->internal;
1279 */
1280 }
1282 /*
1283 * We reached a leaf but did not find the key we were looking for.
1284 * If this is an insert we will be properly positioned for an insert
1285 * (ENOENT) or spike (ENOSPC) operation.
1286 */
1288 done:
1290 }
1292 /*
1293 * Heuristical search for the first element whos comparison is <= 1. May
1294 * return an index whos compare result is > 1 but may only return an index
1295 * whos compare result is <= 1 if it is the first element with that result.
1296 */
1297 int
1299 {
1305 /*
1306 * Don't bother if the node does not have very many elements
1307 */
1318 }
1319 }
1321 }
1324 /************************************************************************
1325 * SPLITTING AND MERGING *
1326 ************************************************************************
1327 *
1328 * These routines do all the dirty work required to split and merge nodes.
1329 */
1331 /*
1332 * Split an internal node into two nodes and move the separator at the split
1333 * point to the parent.
1334 *
1335 * (cursor->node, cursor->index) indicates the element the caller intends
1336 * to push into. We will adjust node and index if that element winds
1337 * up in the split node.
1338 *
1339 * If we are at the root of the filesystem a new root must be created with
1340 * two elements, one pointing to the original root and one pointing to the
1341 * newly allocated split node.
1342 */
1343 static
1344 int
1346 {
1347 hammer_node_ondisk_t ondisk;
1348 hammer_node_t node;
1349 hammer_node_t parent;
1350 hammer_node_t new_node;
1351 hammer_btree_elm_t elm;
1352 hammer_btree_elm_t parent_elm;
1369 /*
1370 * We are splitting but elms[split] will be promoted to the parent,
1371 * leaving the right hand node with one less element. If the
1372 * insertion point will be on the left-hand side adjust the split
1373 * point to give the right hand side one additional node.
1374 */
1381 /*
1382 * If we are at the root of the filesystem, create a new root node
1383 * with 1 element and split normally. Avoid making major
1384 * modifications until we know the whole operation will work.
1385 */
1401 /* ondisk->elms[1].base.btype - not used */
1408 }
1410 /*
1411 * Split node into new_node at the split point.
1412 *
1413 * B O O O P N N B <-- P = node->elms[split]
1414 * 0 1 2 3 4 5 6 <-- subtree indices
1415 *
1416 * x x P x x
1417 * s S S s
1418 * / \
1419 * B O O O B B N N B <--- inner boundary points are 'P'
1420 * 0 1 2 3 4 5 6
1421 *
1422 */
1429 }
1431 }
1434 /*
1435 * Create the new node. P becomes the left-hand boundary in the
1436 * new node. Copy the right-hand boundary as well.
1437 *
1438 * elm is the new separator.
1439 */
1451 /*
1452 * Cleanup the original node. Elm (P) becomes the new boundary,
1453 * its subtree_offset was moved to the new node. If we had created
1454 * a new root its parent pointer may have changed.
1455 */
1459 /*
1460 * Insert the separator into the parent, fixup the parent's
1461 * reference to the original node, and reference the new node.
1462 * The separator is P.
1463 *
1464 * Remember that base.count does not include the right-hand boundary.
1465 */
1478 /*
1479 * The children of new_node need their parent pointer set to new_node.
1480 * The children have already been locked by
1481 * hammer_btree_lock_children().
1482 */
1488 }
1489 }
1492 /*
1493 * The filesystem's root B-Tree pointer may have to be updated.
1494 */
1496 hammer_volume_t volume;
1502 vol0_btree_root);
1509 }
1512 }
1516 /*
1517 * Ok, now adjust the cursor depending on which element the original
1518 * index was pointing at. If we are >= the split point the push node
1519 * is now in the new node.
1520 *
1521 * NOTE: If we are at the split point itself we cannot stay with the
1522 * original node because the push index will point at the right-hand
1523 * boundary, which is illegal.
1524 *
1525 * NOTE: The cursor's parent or parent_index must be adjusted for
1526 * the case where a new parent (new root) was created, and the case
1527 * where the cursor is now pointing at the split node.
1528 */
1539 }
1541 /*
1542 * Fixup left and right bounds
1543 */
1552 done:
1556 }
1558 /*
1559 * Same as the above, but splits a full leaf node.
1560 *
1561 * This function
1562 */
1563 static
1564 int
1566 {
1567 hammer_node_ondisk_t ondisk;
1568 hammer_node_t parent;
1569 hammer_node_t leaf;
1570 hammer_mount_t hmp;
1571 hammer_node_t new_leaf;
1572 hammer_btree_elm_t elm;
1573 hammer_btree_elm_t parent_elm;
1574 hammer_base_elm_t mid_boundary;
1590 /*
1591 * Calculate the split point. If the insertion point will be on
1592 * the left-hand side adjust the split point to give the right
1593 * hand side one additional node.
1594 *
1595 * Spikes are made up of two leaf elements which cannot be
1596 * safely split.
1597 */
1613 /*
1614 * If we are at the root of the tree, create a new root node with
1615 * 1 element and split normally. Avoid making major modifications
1616 * until we know the whole operation will work.
1617 */
1632 /* ondisk->elms[1].base.btype = not used */
1640 }
1642 /*
1643 * Split leaf into new_leaf at the split point. Select a separator
1644 * value in-between the two leafs but with a bent towards the right
1645 * leaf since comparisons use an 'elm >= separator' inequality.
1646 *
1647 * L L L L L L L L
1648 *
1649 * x x P x x
1650 * s S S s
1651 * / \
1652 * L L L L L L L L
1653 */
1660 }
1662 }
1665 /*
1666 * Create the new node and copy the leaf elements from the split
1667 * point on to the new node.
1668 */
1680 /*
1681 * Cleanup the original node. Because this is a leaf node and
1682 * leaf nodes do not have a right-hand boundary, there
1683 * aren't any special edge cases to clean up. We just fixup the
1684 * count.
1685 */
1688 /*
1689 * Insert the separator into the parent, fixup the parent's
1690 * reference to the original node, and reference the new node.
1691 * The separator is P.
1692 *
1693 * Remember that base.count does not include the right-hand boundary.
1694 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1695 */
1711 /*
1712 * The filesystem's root B-Tree pointer may have to be updated.
1713 */
1715 hammer_volume_t volume;
1721 vol0_btree_root);
1728 }
1731 }
1734 /*
1735 * Ok, now adjust the cursor depending on which element the original
1736 * index was pointing at. If we are >= the split point the push node
1737 * is now in the new node.
1738 *
1739 * NOTE: If we are at the split point itself we need to select the
1740 * old or new node based on where key_beg's insertion point will be.
1741 * If we pick the wrong side the inserted element will wind up in
1742 * the wrong leaf node and outside that node's bounds.
1743 */
1756 }
1758 /*
1759 * Fixup left and right bounds
1760 */
1765 /*
1766 * Assert that the bounds are correct.
1767 */
1775 done:
1778 }
1780 /*
1781 * Recursively correct the right-hand boundary's create_tid to (tid) as
1782 * long as the rest of the key matches. We have to recurse upward in
1783 * the tree as well as down the left side of each parent's right node.
1784 *
1785 * Return EDEADLK if we were only partially successful, forcing the caller
1786 * to try again. The original cursor is not modified. This routine can
1787 * also fail with EDEADLK if it is forced to throw away a portion of its
1788 * record history.
1789 *
1790 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1791 */
1794 hammer_node_t node;
1796 };
1800 int
1802 {
1804 hammer_base_elm_t elm;
1805 hammer_node_t orig_node;
1812 /*
1813 * Save our position so we can restore it on return. This also
1814 * gives us a stable 'elm'.
1815 */
1822 /*
1823 * Now build a list of parents going up, allocating a rhb
1824 * structure for each one.
1825 */
1827 /*
1828 * Stop if we no longer have any right-bounds to fix up
1829 */
1834 }
1836 /*
1837 * Stop if the right-hand bound's create_tid does not
1838 * need to be corrected.
1839 */
1851 }
1853 /*
1854 * now safely adjust the right hand bound for each rhb. This may
1855 * also require taking the right side of the tree and iterating down
1856 * ITS left side.
1857 */
1870 /*
1871 * Right-boundary for parent at internal node
1872 * is one element to the right of the element whos
1873 * right boundary needs adjusting. We must then
1874 * traverse down the left side correcting any left
1875 * bounds (which may now be too far to the left).
1876 */
1884 }
1885 }
1887 /*
1888 * Cleanup
1889 */
1895 }
1900 }
1902 /*
1903 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1904 * bound going downward starting at the current cursor position.
1905 *
1906 * This function does not restore the cursor after use.
1907 */
1908 int
1910 {
1912 hammer_base_elm_t elm;
1913 hammer_base_elm_t cmp;
1921 /*
1922 * Record the node and traverse down the left-hand side for all
1923 * matching records needing a boundary correction.
1924 */
1935 /*
1936 * Nothing to traverse down if we are at the right
1937 * boundary of an internal node.
1938 */
1946 }
1956 }
1960 }
1962 /*
1963 * Now we can safely adjust the left-hand boundary from the bottom-up.
1964 * The last element we remove from the list is the caller's right hand
1965 * boundary, which must also be adjusted.
1966 */
1985 }
1986 }
1988 /*
1989 * Cleanup
1990 */
1996 }
1998 }
2000 /*
2001 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2002 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2003 * the operation due to a deadlock, or some other error.
2004 *
2005 * This routine is always called with an empty, locked leaf but may recurse
2006 * into want-to-be-empty parents as part of its operation.
2007 *
2008 * It should also be noted that when removing empty leaves we must be sure
2009 * to test and update mirror_tid because another thread may have deadlocked
2010 * against us (or someone) trying to propogate it up and cannot retry once
2011 * the node has been deleted.
2012 *
2013 * On return the cursor may end up pointing to an internal node, suitable
2014 * for further iteration but not for an immediate insertion or deletion.
2015 */
2016 static int
2018 {
2019 hammer_node_ondisk_t ondisk;
2020 hammer_btree_elm_t elm;
2021 hammer_node_t node;
2022 hammer_node_t parent;
2028 /*
2029 * When deleting the root of the filesystem convert it to
2030 * an empty leaf node. Internal nodes cannot be empty.
2031 */
2042 }
2046 /*
2047 * If another thread deadlocked trying to propogate mirror_tid up
2048 * we have to finish the job before deleting node. XXX
2049 */
2053 parent,
2057 }
2059 /*
2060 * Attempt to remove the parent's reference to the child. If the
2061 * parent would become empty we have to recurse. If we fail we
2062 * leave the parent pointing to an empty leaf node.
2063 */
2065 /*
2066 * This special cursor_up_locked() call leaves the original
2067 * node exclusively locked and referenced, leaves the
2068 * original parent locked (as the new node), and locks the
2069 * new parent. It can return EDEADLK.
2070 */
2086 }
2093 }
2097 /*
2098 * Delete the subtree reference in the parent
2099 */
2114 /*
2115 * cursor->node is invalid, cursor up to make the cursor
2116 * valid again.
2117 */
2119 }
2121 }
2123 /*
2124 * Propagate a mirror TID update upwards through the B-Tree to the root.
2125 *
2126 * A locked internal node must be passed in. The node will remain locked
2127 * on return.
2128 *
2129 * This function syncs mirror_tid at the specified internal node's element,
2130 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2131 */
2132 int
2135 {
2136 hammer_btree_internal_elm_t elm;
2137 hammer_node_t parent;
2143 /*
2144 * Adjust the node's element
2145 */
2154 /*
2155 * Adjust the node's mirror_tid aggragator
2156 */
2174 }
2175 }
2177 }
2179 hammer_node_t
2182 {
2183 hammer_node_t parent;
2184 hammer_btree_elm_t elm;
2187 /*
2188 * Get the node
2189 */
2194 }
2197 /*
2198 * Lock the node
2199 */
2205 }
2208 }
2210 /*
2211 * Figure out which element in the parent is pointing to the
2212 * child.
2213 */
2219 }
2225 }
2229 }
2233 }
2235 /*
2236 * The element (elm) has been moved to a new internal node (node).
2237 *
2238 * If the element represents a pointer to an internal node that node's
2239 * parent must be adjusted to the element's new location.
2240 *
2241 * XXX deadlock potential here with our exclusive locks
2242 */
2243 int
2245 hammer_btree_elm_t elm)
2246 {
2247 hammer_node_t child;
2262 }
2266 }
2268 }
2270 /*
2271 * Exclusively lock all the children of node. This is used by the split
2272 * code to prevent anyone from accessing the children of a cursor node
2273 * while we fix-up its parent offset.
2274 *
2275 * If we don't lock the children we can really mess up cursors which block
2276 * trying to cursor-up into our node.
2277 *
2278 * On failure EDEADLK (or some other error) is returned. If a deadlock
2279 * error is returned the cursor is adjusted to block on termination.
2280 */
2281 int
2284 {
2285 hammer_node_t node;
2286 hammer_node_locklist_t item;
2287 hammer_node_ondisk_t ondisk;
2288 hammer_btree_elm_t elm;
2289 hammer_node_t child;
2297 /*
2298 * We really do not want to block on I/O with exclusive locks held,
2299 * pre-get the children before trying to lock the mess.
2300 */
2307 }
2313 }
2315 /*
2316 * Do it for real
2317 */
2333 }
2339 }
2348 }
2349 }
2350 }
2354 }
2357 /*
2358 * Release previously obtained node locks.
2359 */
2360 void
2362 {
2363 hammer_node_locklist_t item;
2370 }
2371 }
2373 /************************************************************************
2374 * MISCELLANIOUS SUPPORT *
2375 ************************************************************************/
2377 /*
2378 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2379 *
2380 * Note that for this particular function a return value of -1, 0, or +1
2381 * can denote a match if create_tid is otherwise discounted. A create_tid
2382 * of zero is considered to be 'infinity' in comparisons.
2383 *
2384 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2385 */
2386 int
2388 {
2409 /*
2410 * A create_tid of zero indicates a record which is undeletable
2411 * and must be considered to have a value of positive infinity.
2412 */
2417 }
2425 }
2427 /*
2428 * Test a timestamp against an element to determine whether the
2429 * element is visible. A timestamp of 0 means 'infinity'.
2430 */
2431 int
2433 {
2438 }
2444 }
2446 /*
2447 * Create a separator half way inbetween key1 and key2. For fields just
2448 * one unit apart, the separator will match key2. key1 is on the left-hand
2449 * side and key2 is on the right-hand side.
2450 *
2451 * key2 must be >= the separator. It is ok for the separator to match key2.
2452 *
2453 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2454 * key2.
2455 *
2456 * NOTE: It might be beneficial to just scrap this whole mess and just
2457 * set the separator to key2.
2458 */
2459 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2460 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2462 static void
2464 hammer_base_elm_t dest)
2465 {
2480 /*
2481 * Don't bother creating a separator for
2482 * create_tid, which also conveniently avoids
2483 * having to handle the create_tid == 0
2484 * (infinity) case. Just leave create_tid
2485 * set to key2.
2486 *
2487 * Worst case, dest matches key2 exactly,
2488 * which is acceptable.
2489 */
2490 }
2491 }
2492 }
2493 }
2495 #undef MAKE_SEPARATOR
2497 /*
2498 * Return whether a generic internal or leaf node is full
2499 */
2500 static int
2502 {
2514 }
2516 }
2518 #if 0
2519 static int
2521 {
2527 }
2528 #endif
2530 void
2532 {
2533 hammer_btree_elm_t elm;
2539 /*
2540 * Dump both boundary elements if an internal node
2541 */
2546 }
2551 }
2552 }
2553 }
2555 void
2557 {
2580 }
2581 }