| blob | f3fe3ad45c75f8fc82ce19d1dd4199caddf841b1 |
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 1.76 2008/08/06 15:38:58 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>
91 hammer_tid_t mirror_tid);
96 /*
97 * Iterate records after a search. The cursor is iterated forwards past
98 * the current record until a record matching the key-range requirements
99 * is found. ENOENT is returned if the iteration goes past the ending
100 * key.
101 *
102 * The iteration is inclusive of key_beg and can be inclusive or exclusive
103 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
104 *
105 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
106 * may be modified by B-Tree functions.
107 *
108 * cursor->key_beg may or may not be modified by this function during
109 * the iteration. XXX future - in case of an inverted lock we may have
110 * to reinitiate the lookup and set key_beg to properly pick up where we
111 * left off.
112 *
113 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
114 */
115 int
117 {
118 hammer_node_ondisk_t node;
119 hammer_btree_elm_t elm;
124 /*
125 * Skip past the current record
126 */
133 }
135 /*
136 * Loop until an element is found or we are done.
137 */
139 /*
140 * We iterate up the tree and then index over one element
141 * while we are at the last element in the current node.
142 *
143 * If we are at the root of the filesystem, cursor_up
144 * returns ENOENT.
145 *
146 * XXX this could be optimized by storing the information in
147 * the parent reference.
148 *
149 * XXX we can lose the node lock temporarily, this could mess
150 * up our scan.
151 */
162 curthread);
163 }
164 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
168 /* reload stale pointer */
172 /*
173 * If we are reblocking we want to return internal
174 * nodes. Note that the internal node will be
175 * returned multiple times, on each upward recursion
176 * from its children. The caller selects which
177 * revisit it cares about (usually first or last only).
178 */
182 }
185 }
187 /*
188 * Check internal or leaf element. Determine if the record
189 * at the cursor has gone beyond the end of our range.
190 *
191 * We recurse down through internal nodes.
192 */
206 r,
207 curthread
208 );
216 s
217 );
218 }
223 }
228 }
231 /*
232 * Better not be zero
233 */
236 /*
237 * If running the mirror filter see if we can skip
238 * one or more entire sub-trees. If we can we
239 * return the internal mode and the caller processes
240 * the skipped range (see mirror_read)
241 */
247 }
248 }
254 /* reload stale pointer */
270 r
271 );
272 }
276 }
278 /*
279 * We support both end-inclusive and
280 * end-exclusive searches.
281 */
286 }
294 }
300 }
303 }
304 /*
305 * node pointer invalid after loop
306 */
308 /*
309 * Return entry
310 */
320 );
321 }
323 }
325 }
327 /*
328 * We hit an internal element that we could skip as part of a mirroring
329 * scan. Calculate the entire range being skipped.
330 *
331 * It is important to include any gaps between the parent's left_bound
332 * and the node's left_bound, and same goes for the right side.
333 */
334 static void
336 {
338 hammer_node_ondisk_t ondisk;
339 hammer_btree_elm_t elm;
344 /*
345 * Calculate the skipped range
346 */
350 else
357 }
360 else
363 /*
364 * clip the returned result.
365 */
370 }
372 /*
373 * Iterate in the reverse direction. This is used by the pruning code to
374 * avoid overlapping records.
375 */
376 int
378 {
379 hammer_node_ondisk_t node;
380 hammer_btree_elm_t elm;
385 /* mirror filtering not supported for reverse iteration */
388 /*
389 * Skip past the current record. For various reasons the cursor
390 * may end up set to -1 or set to point at the end of the current
391 * node. These cases must be addressed.
392 */
399 }
403 /*
404 * Loop until an element is found or we are done.
405 */
410 /*
411 * We iterate up the tree and then index over one element
412 * while we are at the last element in the current node.
413 */
419 }
420 /* reload stale pointer */
425 }
427 /*
428 * Check internal or leaf element. Determine if the record
429 * at the cursor has gone beyond the end of our range.
430 *
431 * We recurse down through internal nodes.
432 */
446 r
447 );
455 s
456 );
457 }
462 }
465 /*
466 * Better not be zero
467 */
474 /* reload stale pointer */
477 /* this can assign -1 if the leaf was empty */
493 s
494 );
495 }
499 }
507 }
513 }
516 }
517 /*
518 * node pointer invalid after loop
519 */
521 /*
522 * Return entry
523 */
533 );
534 }
536 }
538 }
540 /*
541 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
542 * could not be found, EDEADLK if inserting and a retry is needed, and a
543 * fatal error otherwise. When retrying, the caller must terminate the
544 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
545 *
546 * The cursor is suitably positioned for a deletion on success, and suitably
547 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
548 * specified.
549 *
550 * The cursor may begin anywhere, the search will traverse the tree in
551 * either direction to locate the requested element.
552 *
553 * Most of the logic implementing historical searches is handled here. We
554 * do an initial lookup with create_tid set to the asof TID. Due to the
555 * way records are laid out, a backwards iteration may be required if
556 * ENOENT is returned to locate the historical record. Here's the
557 * problem:
558 *
559 * create_tid: 10 15 20
560 * LEAF1 LEAF2
561 * records: (11) (18)
562 *
563 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
564 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
565 * not visible and thus causes ENOENT to be returned. We really need
566 * to check record 11 in LEAF1. If it also fails then the search fails
567 * (e.g. it might represent the range 11-16 and thus still not match our
568 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
569 * further iterations.
570 *
571 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
572 * and the cursor->create_check TID if an iteration might be needed.
573 * In the above example create_check would be set to 14.
574 */
575 int
577 {
591 /*
592 * Stop if no error.
593 * Stop if error other then ENOENT.
594 * Stop if ENOENT and not special case.
595 */
597 }
601 }
603 /* loop */
604 }
607 }
611 }
613 /*
614 * Execute the logic required to start an iteration. The first record
615 * located within the specified range is returned and iteration control
616 * flags are adjusted for successive hammer_btree_iterate() calls.
617 *
618 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
619 * in a loop without worrying about it. Higher-level merged searches will
620 * adjust the flag appropriately.
621 */
622 int
624 {
631 }
634 }
636 /*
637 * Similarly but for an iteration in the reverse direction.
638 *
639 * Set ATEDISK when iterating backwards to skip the current entry,
640 * which after an ENOENT lookup will be pointing beyond our end point.
641 *
642 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
643 * in a loop without worrying about it. Higher-level merged searches will
644 * adjust the flag appropriately.
645 */
646 int
648 {
660 }
663 }
665 /*
666 * Extract the record and/or data associated with the cursor's current
667 * position. Any prior record or data stored in the cursor is replaced.
668 * The cursor must be positioned at a leaf node.
669 *
670 * NOTE: All extractions occur at the leaf of the B-Tree.
671 */
672 int
674 {
675 hammer_node_ondisk_t node;
676 hammer_btree_elm_t elm;
677 hammer_off_t data_off;
678 hammer_mount_t hmp;
682 /*
683 * The case where the data reference resolves to the same buffer
684 * as the record reference must be handled.
685 */
691 /*
692 * There is nothing to extract for an internal element.
693 */
697 /*
698 * Only record types have data.
699 */
712 /*
713 * Load the data
714 */
725 else
727 }
729 }
732 /*
733 * Insert a leaf element into the B-Tree at the current cursor position.
734 * The cursor is positioned such that the element at and beyond the cursor
735 * are shifted to make room for the new record.
736 *
737 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
738 * flag set and that call must return ENOENT before this function can be
739 * called.
740 *
741 * The caller may depend on the cursor's exclusive lock after return to
742 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
743 *
744 * ENOSPC is returned if there is no room to insert a new record.
745 */
746 int
749 {
750 hammer_node_ondisk_t node;
759 /*
760 * Insert the element at the leaf node and update the count in the
761 * parent. It is possible for parent to be NULL, indicating that
762 * the filesystem's ROOT B-Tree node is a leaf itself, which is
763 * possible. The root inode can never be deleted so the leaf should
764 * never be empty.
765 *
766 * Remember that the right-hand boundary is not included in the
767 * count.
768 */
778 }
783 /*
784 * Update the leaf node's aggregate mirror_tid for mirroring
785 * support.
786 */
790 }
794 }
797 /*
798 * Debugging sanity checks.
799 */
804 }
809 }
811 /*
812 * Delete a record from the B-Tree at the current cursor position.
813 * The cursor is positioned such that the current element is the one
814 * to be deleted.
815 *
816 * On return the cursor will be positioned after the deleted element and
817 * MAY point to an internal node. It will be suitable for the continuation
818 * of an iteration but not for an insertion or deletion.
819 *
820 * Deletions will attempt to partially rebalance the B-Tree in an upward
821 * direction, but will terminate rather then deadlock. Empty internal nodes
822 * are never allowed by a deletion which deadlocks may end up giving us an
823 * empty leaf. The pruner will clean up and rebalance the tree.
824 *
825 * This function can return EDEADLK, requiring the caller to retry the
826 * operation after clearing the deadlock.
827 */
828 int
830 {
831 hammer_node_ondisk_t ondisk;
832 hammer_node_t node;
833 hammer_node_t parent;
842 /*
843 * Delete the element from the leaf node.
844 *
845 * Remember that leaf nodes do not have boundaries.
846 */
857 }
862 /*
863 * Validate local parent
864 */
870 }
872 /*
873 * If the leaf becomes empty it must be detached from the parent,
874 * potentially recursing through to the filesystem root.
875 *
876 * This may reposition the cursor at one of the parent's of the
877 * current node.
878 *
879 * Ignore deadlock errors, that simply means that btree_remove
880 * was unable to recurse and had to leave us with an empty leaf.
881 */
889 }
893 }
895 /*
896 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
897 *
898 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
899 *
900 * The search can begin ANYWHERE in the B-Tree. As a first step the search
901 * iterates up the tree as necessary to properly position itself prior to
902 * actually doing the sarch.
903 *
904 * INSERTIONS: The search will split full nodes and leaves on its way down
905 * and guarentee that the leaf it ends up on is not full. If we run out
906 * of space the search continues to the leaf (to position the cursor for
907 * the spike), but ENOSPC is returned.
908 *
909 * The search is only guarenteed to end up on a leaf if an error code of 0
910 * is returned, or if inserting and an error code of ENOENT is returned.
911 * Otherwise it can stop at an internal node. On success a search returns
912 * a leaf node.
913 *
914 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
915 * filesystem, and it is not simple code. Please note the following facts:
916 *
917 * - Internal node recursions have a boundary on the left AND right. The
918 * right boundary is non-inclusive. The create_tid is a generic part
919 * of the key for internal nodes.
920 *
921 * - Leaf nodes contain terminal elements only now.
922 *
923 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
924 * historical search. ASOF and INSERT are mutually exclusive. When
925 * doing an as-of lookup btree_search() checks for a right-edge boundary
926 * case. If while recursing down the left-edge differs from the key
927 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
928 * with cursor->create_check. This is used by btree_lookup() to iterate.
929 * The iteration backwards because as-of searches can wind up going
930 * down the wrong branch of the B-Tree.
931 */
932 static
933 int
935 {
936 hammer_node_ondisk_t node;
937 hammer_btree_elm_t elm;
956 curthread
957 );
969 );
970 }
972 /*
973 * Move our cursor up the tree until we find a node whos range covers
974 * the key we are trying to locate.
975 *
976 * The left bound is inclusive, the right bound is non-inclusive.
977 * It is ok to cursor up too far.
978 */
989 }
991 /*
992 * The delete-checks below are based on node, not parent. Set the
993 * initial delete-check based on the parent.
994 */
999 }
1001 /*
1002 * We better have ended up with a node somewhere.
1003 */
1006 /*
1007 * If we are inserting we can't start at a full node if the parent
1008 * is also full (because there is no way to split the node),
1009 * continue running up the tree until the requirement is satisfied
1010 * or we hit the root of the filesystem.
1011 *
1012 * (If inserting we aren't doing an as-of search so we don't have
1013 * to worry about create_check).
1014 */
1022 }
1026 }
1029 /* node may have become stale */
1032 }
1034 /*
1035 * Push down through internal nodes to locate the requested key.
1036 */
1039 /*
1040 * Scan the node to find the subtree index to push down into.
1041 * We go one-past, then back-up.
1042 *
1043 * We must proactively remove deleted elements which may
1044 * have been left over from a deadlocked btree_remove().
1045 *
1046 * The left and right boundaries are included in the loop
1047 * in order to detect edge cases.
1048 *
1049 * If the separator only differs by create_tid (r == 1)
1050 * and we are doing an as-of search, we may end up going
1051 * down a branch to the left of the one containing the
1052 * desired key. This requires numerous special cases.
1053 */
1059 }
1061 /*
1062 * Try to shortcut the search before dropping into the
1063 * linear loop. Locate the first node where r <= 1.
1064 */
1073 }
1080 }
1082 }
1086 }
1088 /*
1089 * These cases occur when the parent's idea of the boundary
1090 * is wider then the child's idea of the boundary, and
1091 * require special handling. If not inserting we can
1092 * terminate the search early for these cases but the
1093 * child's boundaries cannot be unconditionally modified.
1094 */
1096 /*
1097 * If i == 0 the search terminated to the LEFT of the
1098 * left_boundary but to the RIGHT of the parent's left
1099 * boundary.
1100 */
1101 u_int8_t save;
1105 /*
1106 * If we aren't inserting we can stop here.
1107 */
1112 }
1114 /*
1115 * Correct a left-hand boundary mismatch.
1116 *
1117 * We can only do this if we can upgrade the lock,
1118 * and synchronized as a background cursor (i.e.
1119 * inserting or pruning).
1120 *
1121 * WARNING: We can only do this if inserting, i.e.
1122 * we are running on the backend.
1123 */
1134 /*
1135 * If i == node->count + 1 the search terminated to
1136 * the RIGHT of the right boundary but to the LEFT
1137 * of the parent's right boundary. If we aren't
1138 * inserting we can stop here.
1139 *
1140 * Note that the last element in this case is
1141 * elms[i-2] prior to adjustments to 'i'.
1142 */
1148 }
1150 /*
1151 * Correct a right-hand boundary mismatch.
1152 * (actual push-down record is i-2 prior to
1153 * adjustments to i).
1154 *
1155 * We can only do this if we can upgrade the lock,
1156 * and synchronized as a background cursor (i.e.
1157 * inserting or pruning).
1158 *
1159 * WARNING: We can only do this if inserting, i.e.
1160 * we are running on the backend.
1161 */
1172 /*
1173 * The push-down index is now i - 1. If we had
1174 * terminated on the right boundary this will point
1175 * us at the last element.
1176 */
1178 }
1186 i,
1192 );
1193 }
1195 /*
1196 * We better have a valid subtree offset.
1197 */
1200 /*
1201 * Handle insertion and deletion requirements.
1202 *
1203 * If inserting split full nodes. The split code will
1204 * adjust cursor->node and cursor->index if the current
1205 * index winds up in the new node.
1206 *
1207 * If inserting and a left or right edge case was detected,
1208 * we cannot correct the left or right boundary and must
1209 * prepend and append an empty leaf node in order to make
1210 * the boundary correction.
1211 *
1212 * If we run out of space we set enospc and continue on
1213 * to a leaf to provide the spike code with a good point
1214 * of entry.
1215 */
1223 }
1224 /*
1225 * reload stale pointers
1226 */
1229 }
1230 }
1232 /*
1233 * Push down (push into new node, existing node becomes
1234 * the parent) and continue the search.
1235 */
1237 /* node may have become stale */
1241 }
1243 /*
1244 * We are at a leaf, do a linear search of the key array.
1245 *
1246 * On success the index is set to the matching element and 0
1247 * is returned.
1248 *
1249 * On failure the index is set to the insertion point and ENOENT
1250 * is returned.
1251 *
1252 * Boundaries are not stored in leaf nodes, so the index can wind
1253 * up to the left of element 0 (index == 0) or past the end of
1254 * the array (index == node->count). It is also possible that the
1255 * leaf might be empty.
1256 */
1264 }
1266 /*
1267 * Try to shortcut the search before dropping into the
1268 * linear loop. Locate the first node where r <= 1.
1269 */
1280 /*
1281 * We are at a record element. Stop if we've flipped past
1282 * key_beg, not counting the create_tid test. Allow the
1283 * r == 1 case (key_beg > element but differs only by its
1284 * create_tid) to fall through to the AS-OF check.
1285 */
1293 }
1295 /*
1296 * Check our as-of timestamp against the element.
1297 */
1303 }
1304 /* success */
1309 }
1310 /* success */
1311 }
1317 }
1319 }
1321 /*
1322 * The search of the leaf node failed. i is the insertion point.
1323 */
1324 failed:
1328 }
1330 /*
1331 * No exact match was found, i is now at the insertion point.
1332 *
1333 * If inserting split a full leaf before returning. This
1334 * may have the side effect of adjusting cursor->node and
1335 * cursor->index.
1336 */
1345 }
1346 /*
1347 * reload stale pointers
1348 */
1349 /* NOT USED
1350 i = cursor->index;
1351 node = &cursor->node->internal;
1352 */
1353 }
1355 /*
1356 * We reached a leaf but did not find the key we were looking for.
1357 * If this is an insert we will be properly positioned for an insert
1358 * (ENOENT) or spike (ENOSPC) operation.
1359 */
1361 done:
1363 }
1365 /*
1366 * Heuristical search for the first element whos comparison is <= 1. May
1367 * return an index whos compare result is > 1 but may only return an index
1368 * whos compare result is <= 1 if it is the first element with that result.
1369 */
1370 int
1372 {
1378 /*
1379 * Don't bother if the node does not have very many elements
1380 */
1391 }
1392 }
1394 }
1397 /************************************************************************
1398 * SPLITTING AND MERGING *
1399 ************************************************************************
1400 *
1401 * These routines do all the dirty work required to split and merge nodes.
1402 */
1404 /*
1405 * Split an internal node into two nodes and move the separator at the split
1406 * point to the parent.
1407 *
1408 * (cursor->node, cursor->index) indicates the element the caller intends
1409 * to push into. We will adjust node and index if that element winds
1410 * up in the split node.
1411 *
1412 * If we are at the root of the filesystem a new root must be created with
1413 * two elements, one pointing to the original root and one pointing to the
1414 * newly allocated split node.
1415 */
1416 static
1417 int
1419 {
1420 hammer_node_ondisk_t ondisk;
1421 hammer_node_t node;
1422 hammer_node_t parent;
1423 hammer_node_t new_node;
1424 hammer_btree_elm_t elm;
1425 hammer_btree_elm_t parent_elm;
1428 hammer_off_t hint;
1444 /*
1445 * We are splitting but elms[split] will be promoted to the parent,
1446 * leaving the right hand node with one less element. If the
1447 * insertion point will be on the left-hand side adjust the split
1448 * point to give the right hand side one additional node.
1449 */
1456 /*
1457 * If we are at the root of the filesystem, create a new root node
1458 * with 1 element and split normally. Avoid making major
1459 * modifications until we know the whole operation will work.
1460 */
1478 /* ondisk->elms[1].base.btype - not used */
1485 }
1487 /*
1488 * Calculate a hint for the allocation of the new B-Tree node.
1489 * The most likely expansion is coming from the insertion point
1490 * at cursor->index, so try to localize the allocation of our
1491 * new node to accomodate that sub-tree.
1492 *
1493 * Use the right-most sub-tree when expandinging on the right edge.
1494 * This is a very common case when copying a directory tree.
1495 */
1498 else
1501 /*
1502 * Split node into new_node at the split point.
1503 *
1504 * B O O O P N N B <-- P = node->elms[split] (index 4)
1505 * 0 1 2 3 4 5 6 <-- subtree indices
1506 *
1507 * x x P x x
1508 * s S S s
1509 * / \
1510 * B O O O B B N N B <--- inner boundary points are 'P'
1511 * 0 1 2 3 4 5 6
1512 */
1519 }
1521 }
1524 /*
1525 * Create the new node. P becomes the left-hand boundary in the
1526 * new node. Copy the right-hand boundary as well.
1527 *
1528 * elm is the new separator.
1529 */
1543 /*
1544 * Cleanup the original node. Elm (P) becomes the new boundary,
1545 * its subtree_offset was moved to the new node. If we had created
1546 * a new root its parent pointer may have changed.
1547 */
1551 /*
1552 * Insert the separator into the parent, fixup the parent's
1553 * reference to the original node, and reference the new node.
1554 * The separator is P.
1555 *
1556 * Remember that base.count does not include the right-hand boundary.
1557 */
1572 /*
1573 * The children of new_node need their parent pointer set to new_node.
1574 * The children have already been locked by
1575 * hammer_btree_lock_children().
1576 */
1582 }
1583 }
1586 /*
1587 * The filesystem's root B-Tree pointer may have to be updated.
1588 */
1590 hammer_volume_t volume;
1596 vol0_btree_root);
1603 }
1606 }
1609 /*
1610 * Ok, now adjust the cursor depending on which element the original
1611 * index was pointing at. If we are >= the split point the push node
1612 * is now in the new node.
1613 *
1614 * NOTE: If we are at the split point itself we cannot stay with the
1615 * original node because the push index will point at the right-hand
1616 * boundary, which is illegal.
1617 *
1618 * NOTE: The cursor's parent or parent_index must be adjusted for
1619 * the case where a new parent (new root) was created, and the case
1620 * where the cursor is now pointing at the split node.
1621 */
1632 }
1634 /*
1635 * Fixup left and right bounds
1636 */
1645 done:
1649 }
1651 /*
1652 * Same as the above, but splits a full leaf node.
1653 *
1654 * This function
1655 */
1656 static
1657 int
1659 {
1660 hammer_node_ondisk_t ondisk;
1661 hammer_node_t parent;
1662 hammer_node_t leaf;
1663 hammer_mount_t hmp;
1664 hammer_node_t new_leaf;
1665 hammer_btree_elm_t elm;
1666 hammer_btree_elm_t parent_elm;
1667 hammer_base_elm_t mid_boundary;
1668 hammer_off_t hint;
1684 /*
1685 * Calculate the split point. If the insertion point will be on
1686 * the left-hand side adjust the split point to give the right
1687 * hand side one additional node.
1688 *
1689 * Spikes are made up of two leaf elements which cannot be
1690 * safely split.
1691 */
1707 /*
1708 * If we are at the root of the tree, create a new root node with
1709 * 1 element and split normally. Avoid making major modifications
1710 * until we know the whole operation will work.
1711 */
1728 /* ondisk->elms[1].base.btype = not used */
1736 }
1738 /*
1739 * Calculate a hint for the allocation of the new B-Tree leaf node.
1740 * For now just try to localize it within the same bigblock as
1741 * the current leaf.
1742 *
1743 * If the insertion point is at the end of the leaf we recognize a
1744 * likely append sequence of some sort (data, meta-data, inodes,
1745 * whatever). Set the hint to zero to allocate out of linear space
1746 * instead of trying to completely fill previously hinted space.
1747 *
1748 * This also sets the stage for recursive splits to localize using
1749 * the new space.
1750 */
1754 else
1757 /*
1758 * Split leaf into new_leaf at the split point. Select a separator
1759 * value in-between the two leafs but with a bent towards the right
1760 * leaf since comparisons use an 'elm >= separator' inequality.
1761 *
1762 * L L L L L L L L
1763 *
1764 * x x P x x
1765 * s S S s
1766 * / \
1767 * L L L L L L L L
1768 */
1775 }
1777 }
1780 /*
1781 * Create the new node and copy the leaf elements from the split
1782 * point on to the new node.
1783 */
1797 /*
1798 * Cleanup the original node. Because this is a leaf node and
1799 * leaf nodes do not have a right-hand boundary, there
1800 * aren't any special edge cases to clean up. We just fixup the
1801 * count.
1802 */
1805 /*
1806 * Insert the separator into the parent, fixup the parent's
1807 * reference to the original node, and reference the new node.
1808 * The separator is P.
1809 *
1810 * Remember that base.count does not include the right-hand boundary.
1811 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1812 */
1830 /*
1831 * The filesystem's root B-Tree pointer may have to be updated.
1832 */
1834 hammer_volume_t volume;
1840 vol0_btree_root);
1847 }
1850 }
1853 /*
1854 * Ok, now adjust the cursor depending on which element the original
1855 * index was pointing at. If we are >= the split point the push node
1856 * is now in the new node.
1857 *
1858 * NOTE: If we are at the split point itself we need to select the
1859 * old or new node based on where key_beg's insertion point will be.
1860 * If we pick the wrong side the inserted element will wind up in
1861 * the wrong leaf node and outside that node's bounds.
1862 */
1875 }
1877 /*
1878 * Fixup left and right bounds
1879 */
1884 /*
1885 * Assert that the bounds are correct.
1886 */
1894 done:
1897 }
1899 #if 0
1901 /*
1902 * Recursively correct the right-hand boundary's create_tid to (tid) as
1903 * long as the rest of the key matches. We have to recurse upward in
1904 * the tree as well as down the left side of each parent's right node.
1905 *
1906 * Return EDEADLK if we were only partially successful, forcing the caller
1907 * to try again. The original cursor is not modified. This routine can
1908 * also fail with EDEADLK if it is forced to throw away a portion of its
1909 * record history.
1910 *
1911 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1912 */
1915 hammer_node_t node;
1917 };
1921 int
1923 {
1926 hammer_base_elm_t elm;
1927 hammer_node_t orig_node;
1935 /*
1936 * Save our position so we can restore it on return. This also
1937 * gives us a stable 'elm'.
1938 */
1945 /*
1946 * Now build a list of parents going up, allocating a rhb
1947 * structure for each one.
1948 */
1950 /*
1951 * Stop if we no longer have any right-bounds to fix up
1952 */
1957 }
1959 /*
1960 * Stop if the right-hand bound's create_tid does not
1961 * need to be corrected.
1962 */
1974 }
1976 /*
1977 * now safely adjust the right hand bound for each rhb. This may
1978 * also require taking the right side of the tree and iterating down
1979 * ITS left side.
1980 */
1993 /*
1994 * Right-boundary for parent at internal node
1995 * is one element to the right of the element whos
1996 * right boundary needs adjusting. We must then
1997 * traverse down the left side correcting any left
1998 * bounds (which may now be too far to the left).
1999 */
2007 }
2008 }
2010 /*
2011 * Cleanup
2012 */
2018 }
2023 }
2025 /*
2026 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2027 * bound going downward starting at the current cursor position.
2028 *
2029 * This function does not restore the cursor after use.
2030 */
2031 int
2033 {
2035 hammer_base_elm_t elm;
2036 hammer_base_elm_t cmp;
2046 /*
2047 * Record the node and traverse down the left-hand side for all
2048 * matching records needing a boundary correction.
2049 */
2060 /*
2061 * Nothing to traverse down if we are at the right
2062 * boundary of an internal node.
2063 */
2071 }
2081 }
2085 }
2087 /*
2088 * Now we can safely adjust the left-hand boundary from the bottom-up.
2089 * The last element we remove from the list is the caller's right hand
2090 * boundary, which must also be adjusted.
2091 */
2110 }
2111 }
2113 /*
2114 * Cleanup
2115 */
2121 }
2123 }
2125 #endif
2127 /*
2128 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2129 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2130 * the operation due to a deadlock, or some other error.
2131 *
2132 * This routine is initially called with an empty leaf and may be
2133 * recursively called with single-element internal nodes.
2134 *
2135 * It should also be noted that when removing empty leaves we must be sure
2136 * to test and update mirror_tid because another thread may have deadlocked
2137 * against us (or someone) trying to propagate it up and cannot retry once
2138 * the node has been deleted.
2139 *
2140 * On return the cursor may end up pointing to an internal node, suitable
2141 * for further iteration but not for an immediate insertion or deletion.
2142 */
2143 static int
2145 {
2146 hammer_node_ondisk_t ondisk;
2147 hammer_btree_elm_t elm;
2148 hammer_node_t node;
2149 hammer_node_t parent;
2155 /*
2156 * When deleting the root of the filesystem convert it to
2157 * an empty leaf node. Internal nodes cannot be empty.
2158 */
2169 }
2174 /*
2175 * Attempt to remove the parent's reference to the child. If the
2176 * parent would become empty we have to recurse. If we fail we
2177 * leave the parent pointing to an empty leaf node.
2178 *
2179 * We have to recurse successfully before we can delete the internal
2180 * node as it is illegal to have empty internal nodes. Even though
2181 * the operation may be aborted we must still fixup any unlocked
2182 * cursors as if we had deleted the element prior to recursing
2183 * (by calling hammer_cursor_deleted_element()) so those cursors
2184 * are properly forced up the chain by the recursion.
2185 */
2187 /*
2188 * This special cursor_up_locked() call leaves the original
2189 * node exclusively locked and referenced, leaves the
2190 * original parent locked (as the new node), and locks the
2191 * new parent. It can return EDEADLK.
2192 */
2206 /*
2207 * Defer parent removal because we could not
2208 * get the lock, just let the leaf remain
2209 * empty.
2210 */
2211 /**/
2212 }
2216 /*
2217 * Defer parent removal because we could not
2218 * get the lock, just let the leaf remain
2219 * empty.
2220 */
2221 /**/
2222 }
2234 /*
2235 * We must retain the highest mirror_tid. The deleted
2236 * range is now encompassed by the element to the left.
2237 * If we are already at the left edge the new left edge
2238 * inherits mirror_tid.
2239 *
2240 * Note that bounds of the parent to our parent may create
2241 * a gap to the left of our left-most node or to the right
2242 * of our right-most node. The gap is silently included
2243 * in the mirror_tid's area of effect from the point of view
2244 * of the scan.
2245 */
2251 }
2257 }
2258 }
2260 /*
2261 * Delete the subtree reference in the parent
2262 */
2271 /*
2272 * cursor->node is invalid, cursor up to make the cursor
2273 * valid again.
2274 */
2276 }
2278 }
2280 /*
2281 * Propagate cursor->trans->tid up the B-Tree starting at the current
2282 * cursor position using pseudofs info gleaned from the passed inode.
2283 *
2284 * The passed inode has no relationship to the cursor position other
2285 * then being in the same pseudofs as the insertion or deletion we
2286 * are propagating the mirror_tid for.
2287 */
2288 void
2290 hammer_pseudofs_inmem_t pfsm,
2291 hammer_btree_leaf_elm_t leaf)
2292 {
2293 hammer_cursor_t ncursor;
2294 hammer_tid_t mirror_tid;
2297 /*
2298 * We do not propagate a mirror_tid if the filesystem was mounted
2299 * in no-mirror mode.
2300 */
2304 /*
2305 * This is a bit of a hack because we cannot deadlock or return
2306 * EDEADLK here. The related operation has already completed and
2307 * we must propagate the mirror_tid now regardless.
2308 *
2309 * Generate a new cursor which inherits the original's locks and
2310 * unlock the original. Use the new cursor to propagate the
2311 * mirror_tid. Then clean up the new cursor and reacquire locks
2312 * on the original.
2313 *
2314 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2315 * original's locks and the original is tracked and must be
2316 * re-locked.
2317 */
2324 }
2327 /*
2328 * Propagate a mirror TID update upwards through the B-Tree to the root.
2329 *
2330 * A locked internal node must be passed in. The node will remain locked
2331 * on return.
2332 *
2333 * This function syncs mirror_tid at the specified internal node's element,
2334 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2335 */
2336 static int
2338 {
2339 hammer_btree_internal_elm_t elm;
2340 hammer_node_t node;
2350 }
2356 /*
2357 * Adjust the node's element
2358 */
2370 }
2373 /*
2374 * Adjust the node's mirror_tid aggregator
2375 */
2385 }
2386 }
2390 }
2392 hammer_node_t
2395 {
2396 hammer_node_t parent;
2397 hammer_btree_elm_t elm;
2400 /*
2401 * Get the node
2402 */
2407 }
2410 /*
2411 * Lock the node
2412 */
2418 }
2421 }
2423 /*
2424 * Figure out which element in the parent is pointing to the
2425 * child.
2426 */
2432 }
2438 }
2442 }
2446 }
2448 /*
2449 * The element (elm) has been moved to a new internal node (node).
2450 *
2451 * If the element represents a pointer to an internal node that node's
2452 * parent must be adjusted to the element's new location.
2453 *
2454 * XXX deadlock potential here with our exclusive locks
2455 */
2456 int
2458 hammer_btree_elm_t elm)
2459 {
2460 hammer_node_t child;
2475 }
2479 }
2481 }
2483 /*
2484 * Initialize the root of a recursive B-Tree node lock list structure.
2485 */
2486 void
2488 {
2496 }
2498 /*
2499 * Exclusively lock all the children of node. This is used by the split
2500 * code to prevent anyone from accessing the children of a cursor node
2501 * while we fix-up its parent offset.
2502 *
2503 * If we don't lock the children we can really mess up cursors which block
2504 * trying to cursor-up into our node.
2505 *
2506 * On failure EDEADLK (or some other error) is returned. If a deadlock
2507 * error is returned the cursor is adjusted to block on termination.
2508 *
2509 * The caller is responsible for managing parent->node, the root's node
2510 * is usually aliased from a cursor.
2511 */
2512 int
2514 hammer_node_lock_t parent)
2515 {
2516 hammer_node_t node;
2517 hammer_node_lock_t item;
2518 hammer_node_ondisk_t ondisk;
2519 hammer_btree_elm_t elm;
2520 hammer_node_t child;
2530 /*
2531 * We really do not want to block on I/O with exclusive locks held,
2532 * pre-get the children before trying to lock the mess. This is
2533 * only done one-level deep for now.
2534 */
2541 }
2547 }
2549 /*
2550 * Do it for real
2551 */
2567 }
2573 }
2586 /*
2587 * Recurse (used by the rebalancing code)
2588 */
2591 cursor,
2593 item);
2594 }
2595 }
2596 }
2597 }
2601 }
2603 /*
2604 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2605 * including the parent.
2606 */
2607 void
2609 {
2611 hammer_node_lock_t item;
2615 M_WAITOK);
2617 }
2620 }
2621 }
2623 /*
2624 * Recursively sync modified copies to the media.
2625 */
2626 int
2628 {
2629 hammer_node_lock_t item;
2640 }
2641 }
2644 }
2646 }
2648 /*
2649 * Release previously obtained node locks. The caller is responsible for
2650 * cleaning up parent->node itself (its usually just aliased from a cursor),
2651 * but this function will take care of the copies.
2652 */
2653 void
2655 {
2656 hammer_node_lock_t item;
2661 }
2668 }
2669 }
2671 /************************************************************************
2672 * MISCELLANIOUS SUPPORT *
2673 ************************************************************************/
2675 /*
2676 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2677 *
2678 * Note that for this particular function a return value of -1, 0, or +1
2679 * can denote a match if create_tid is otherwise discounted. A create_tid
2680 * of zero is considered to be 'infinity' in comparisons.
2681 *
2682 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2683 */
2684 int
2686 {
2707 /*
2708 * A create_tid of zero indicates a record which is undeletable
2709 * and must be considered to have a value of positive infinity.
2710 */
2715 }
2723 }
2725 /*
2726 * Test a timestamp against an element to determine whether the
2727 * element is visible. A timestamp of 0 means 'infinity'.
2728 */
2729 int
2731 {
2736 }
2742 }
2744 /*
2745 * Create a separator half way inbetween key1 and key2. For fields just
2746 * one unit apart, the separator will match key2. key1 is on the left-hand
2747 * side and key2 is on the right-hand side.
2748 *
2749 * key2 must be >= the separator. It is ok for the separator to match key2.
2750 *
2751 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2752 * key2.
2753 *
2754 * NOTE: It might be beneficial to just scrap this whole mess and just
2755 * set the separator to key2.
2756 */
2757 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2758 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2760 static void
2762 hammer_base_elm_t dest)
2763 {
2778 /*
2779 * Don't bother creating a separator for
2780 * create_tid, which also conveniently avoids
2781 * having to handle the create_tid == 0
2782 * (infinity) case. Just leave create_tid
2783 * set to key2.
2784 *
2785 * Worst case, dest matches key2 exactly,
2786 * which is acceptable.
2787 */
2788 }
2789 }
2790 }
2791 }
2793 #undef MAKE_SEPARATOR
2795 /*
2796 * Return whether a generic internal or leaf node is full
2797 */
2798 static int
2800 {
2812 }
2814 }
2816 #if 0
2817 static int
2819 {
2825 }
2826 #endif
2828 void
2830 {
2831 hammer_btree_elm_t elm;
2837 /*
2838 * Dump both boundary elements if an internal node
2839 */
2844 }
2849 }
2850 }
2851 }
2853 void
2855 {
2878 }
2879 }