| blob | 2c05e127983d565874d9013115c0acedde5b86a4 |
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;
120 hammer_mount_t hmp;
125 /*
126 * Skip past the current record
127 */
135 }
137 /*
138 * HAMMER can wind up being cpu-bound.
139 */
143 }
146 /*
147 * Loop until an element is found or we are done.
148 */
150 /*
151 * We iterate up the tree and then index over one element
152 * while we are at the last element in the current node.
153 *
154 * If we are at the root of the filesystem, cursor_up
155 * returns ENOENT.
156 *
157 * XXX this could be optimized by storing the information in
158 * the parent reference.
159 *
160 * XXX we can lose the node lock temporarily, this could mess
161 * up our scan.
162 */
173 curthread);
174 }
175 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
179 /* reload stale pointer */
183 /*
184 * If we are reblocking we want to return internal
185 * nodes. Note that the internal node will be
186 * returned multiple times, on each upward recursion
187 * from its children. The caller selects which
188 * revisit it cares about (usually first or last only).
189 */
193 }
196 }
198 /*
199 * Check internal or leaf element. Determine if the record
200 * at the cursor has gone beyond the end of our range.
201 *
202 * We recurse down through internal nodes.
203 */
217 r,
218 curthread
219 );
227 s
228 );
229 }
234 }
239 }
242 /*
243 * Better not be zero
244 */
247 /*
248 * If running the mirror filter see if we can skip
249 * one or more entire sub-trees. If we can we
250 * return the internal mode and the caller processes
251 * the skipped range (see mirror_read)
252 */
258 }
259 }
265 /* reload stale pointer */
281 r
282 );
283 }
287 }
289 /*
290 * We support both end-inclusive and
291 * end-exclusive searches.
292 */
297 }
305 }
311 }
314 }
315 /*
316 * node pointer invalid after loop
317 */
319 /*
320 * Return entry
321 */
331 );
332 }
334 }
336 }
338 /*
339 * We hit an internal element that we could skip as part of a mirroring
340 * scan. Calculate the entire range being skipped.
341 *
342 * It is important to include any gaps between the parent's left_bound
343 * and the node's left_bound, and same goes for the right side.
344 */
345 static void
347 {
349 hammer_node_ondisk_t ondisk;
350 hammer_btree_elm_t elm;
355 /*
356 * Calculate the skipped range
357 */
361 else
368 }
371 else
374 /*
375 * clip the returned result.
376 */
381 }
383 /*
384 * Iterate in the reverse direction. This is used by the pruning code to
385 * avoid overlapping records.
386 */
387 int
389 {
390 hammer_node_ondisk_t node;
391 hammer_btree_elm_t elm;
396 /* mirror filtering not supported for reverse iteration */
399 /*
400 * Skip past the current record. For various reasons the cursor
401 * may end up set to -1 or set to point at the end of the current
402 * node. These cases must be addressed.
403 */
410 }
414 /*
415 * Loop until an element is found or we are done.
416 */
421 /*
422 * We iterate up the tree and then index over one element
423 * while we are at the last element in the current node.
424 */
430 }
431 /* reload stale pointer */
436 }
438 /*
439 * Check internal or leaf element. Determine if the record
440 * at the cursor has gone beyond the end of our range.
441 *
442 * We recurse down through internal nodes.
443 */
457 r
458 );
466 s
467 );
468 }
473 }
476 /*
477 * Better not be zero
478 */
485 /* reload stale pointer */
488 /* this can assign -1 if the leaf was empty */
504 s
505 );
506 }
510 }
518 }
524 }
527 }
528 /*
529 * node pointer invalid after loop
530 */
532 /*
533 * Return entry
534 */
544 );
545 }
547 }
549 }
551 /*
552 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
553 * could not be found, EDEADLK if inserting and a retry is needed, and a
554 * fatal error otherwise. When retrying, the caller must terminate the
555 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
556 *
557 * The cursor is suitably positioned for a deletion on success, and suitably
558 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
559 * specified.
560 *
561 * The cursor may begin anywhere, the search will traverse the tree in
562 * either direction to locate the requested element.
563 *
564 * Most of the logic implementing historical searches is handled here. We
565 * do an initial lookup with create_tid set to the asof TID. Due to the
566 * way records are laid out, a backwards iteration may be required if
567 * ENOENT is returned to locate the historical record. Here's the
568 * problem:
569 *
570 * create_tid: 10 15 20
571 * LEAF1 LEAF2
572 * records: (11) (18)
573 *
574 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
575 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
576 * not visible and thus causes ENOENT to be returned. We really need
577 * to check record 11 in LEAF1. If it also fails then the search fails
578 * (e.g. it might represent the range 11-16 and thus still not match our
579 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
580 * further iterations.
581 *
582 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
583 * and the cursor->create_check TID if an iteration might be needed.
584 * In the above example create_check would be set to 14.
585 */
586 int
588 {
602 /*
603 * Stop if no error.
604 * Stop if error other then ENOENT.
605 * Stop if ENOENT and not special case.
606 */
608 }
612 }
614 /* loop */
615 }
618 }
622 }
624 /*
625 * Execute the logic required to start an iteration. The first record
626 * located within the specified range is returned and iteration control
627 * flags are adjusted for successive hammer_btree_iterate() calls.
628 *
629 * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
630 * in a loop without worrying about it. Higher-level merged searches will
631 * adjust the flag appropriately.
632 */
633 int
635 {
642 }
645 }
647 /*
648 * Similarly but for an iteration in the reverse direction.
649 *
650 * Set ATEDISK when iterating backwards to skip the current entry,
651 * which after an ENOENT lookup will be pointing beyond our end point.
652 *
653 * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
654 * in a loop without worrying about it. Higher-level merged searches will
655 * adjust the flag appropriately.
656 */
657 int
659 {
671 }
674 }
676 /*
677 * Extract the record and/or data associated with the cursor's current
678 * position. Any prior record or data stored in the cursor is replaced.
679 * The cursor must be positioned at a leaf node.
680 *
681 * NOTE: All extractions occur at the leaf of the B-Tree.
682 */
683 int
685 {
686 hammer_node_ondisk_t node;
687 hammer_btree_elm_t elm;
688 hammer_off_t data_off;
689 hammer_mount_t hmp;
693 /*
694 * The case where the data reference resolves to the same buffer
695 * as the record reference must be handled.
696 */
702 /*
703 * There is nothing to extract for an internal element.
704 */
708 /*
709 * Only record types have data.
710 */
723 /*
724 * Load the data
725 */
736 else
738 }
740 }
743 /*
744 * Insert a leaf element into the B-Tree at the current cursor position.
745 * The cursor is positioned such that the element at and beyond the cursor
746 * are shifted to make room for the new record.
747 *
748 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
749 * flag set and that call must return ENOENT before this function can be
750 * called.
751 *
752 * The caller may depend on the cursor's exclusive lock after return to
753 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
754 *
755 * ENOSPC is returned if there is no room to insert a new record.
756 */
757 int
760 {
761 hammer_node_ondisk_t node;
770 /*
771 * Insert the element at the leaf node and update the count in the
772 * parent. It is possible for parent to be NULL, indicating that
773 * the filesystem's ROOT B-Tree node is a leaf itself, which is
774 * possible. The root inode can never be deleted so the leaf should
775 * never be empty.
776 *
777 * Remember that the right-hand boundary is not included in the
778 * count.
779 */
789 }
794 /*
795 * Update the leaf node's aggregate mirror_tid for mirroring
796 * support.
797 */
801 }
805 }
808 /*
809 * Debugging sanity checks.
810 */
815 }
820 }
822 /*
823 * Delete a record from the B-Tree at the current cursor position.
824 * The cursor is positioned such that the current element is the one
825 * to be deleted.
826 *
827 * On return the cursor will be positioned after the deleted element and
828 * MAY point to an internal node. It will be suitable for the continuation
829 * of an iteration but not for an insertion or deletion.
830 *
831 * Deletions will attempt to partially rebalance the B-Tree in an upward
832 * direction, but will terminate rather then deadlock. Empty internal nodes
833 * are never allowed by a deletion which deadlocks may end up giving us an
834 * empty leaf. The pruner will clean up and rebalance the tree.
835 *
836 * This function can return EDEADLK, requiring the caller to retry the
837 * operation after clearing the deadlock.
838 */
839 int
841 {
842 hammer_node_ondisk_t ondisk;
843 hammer_node_t node;
844 hammer_node_t parent;
853 /*
854 * Delete the element from the leaf node.
855 *
856 * Remember that leaf nodes do not have boundaries.
857 */
868 }
873 /*
874 * Validate local parent
875 */
881 }
883 /*
884 * If the leaf becomes empty it must be detached from the parent,
885 * potentially recursing through to the filesystem root.
886 *
887 * This may reposition the cursor at one of the parent's of the
888 * current node.
889 *
890 * Ignore deadlock errors, that simply means that btree_remove
891 * was unable to recurse and had to leave us with an empty leaf.
892 */
900 }
904 }
906 /*
907 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
908 *
909 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
910 *
911 * The search can begin ANYWHERE in the B-Tree. As a first step the search
912 * iterates up the tree as necessary to properly position itself prior to
913 * actually doing the sarch.
914 *
915 * INSERTIONS: The search will split full nodes and leaves on its way down
916 * and guarentee that the leaf it ends up on is not full. If we run out
917 * of space the search continues to the leaf (to position the cursor for
918 * the spike), but ENOSPC is returned.
919 *
920 * The search is only guarenteed to end up on a leaf if an error code of 0
921 * is returned, or if inserting and an error code of ENOENT is returned.
922 * Otherwise it can stop at an internal node. On success a search returns
923 * a leaf node.
924 *
925 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
926 * filesystem, and it is not simple code. Please note the following facts:
927 *
928 * - Internal node recursions have a boundary on the left AND right. The
929 * right boundary is non-inclusive. The create_tid is a generic part
930 * of the key for internal nodes.
931 *
932 * - Leaf nodes contain terminal elements only now.
933 *
934 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
935 * historical search. ASOF and INSERT are mutually exclusive. When
936 * doing an as-of lookup btree_search() checks for a right-edge boundary
937 * case. If while recursing down the left-edge differs from the key
938 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
939 * with cursor->create_check. This is used by btree_lookup() to iterate.
940 * The iteration backwards because as-of searches can wind up going
941 * down the wrong branch of the B-Tree.
942 */
943 static
944 int
946 {
947 hammer_node_ondisk_t node;
948 hammer_btree_elm_t elm;
967 curthread
968 );
981 );
982 }
984 /*
985 * Move our cursor up the tree until we find a node whos range covers
986 * the key we are trying to locate.
987 *
988 * The left bound is inclusive, the right bound is non-inclusive.
989 * It is ok to cursor up too far.
990 */
1001 }
1003 /*
1004 * The delete-checks below are based on node, not parent. Set the
1005 * initial delete-check based on the parent.
1006 */
1011 }
1013 /*
1014 * We better have ended up with a node somewhere.
1015 */
1018 /*
1019 * If we are inserting we can't start at a full node if the parent
1020 * is also full (because there is no way to split the node),
1021 * continue running up the tree until the requirement is satisfied
1022 * or we hit the root of the filesystem.
1023 *
1024 * (If inserting we aren't doing an as-of search so we don't have
1025 * to worry about create_check).
1026 */
1034 }
1038 }
1041 /* node may have become stale */
1044 }
1046 /*
1047 * Push down through internal nodes to locate the requested key.
1048 */
1051 /*
1052 * Scan the node to find the subtree index to push down into.
1053 * We go one-past, then back-up.
1054 *
1055 * We must proactively remove deleted elements which may
1056 * have been left over from a deadlocked btree_remove().
1057 *
1058 * The left and right boundaries are included in the loop
1059 * in order to detect edge cases.
1060 *
1061 * If the separator only differs by create_tid (r == 1)
1062 * and we are doing an as-of search, we may end up going
1063 * down a branch to the left of the one containing the
1064 * desired key. This requires numerous special cases.
1065 */
1071 }
1073 /*
1074 * Try to shortcut the search before dropping into the
1075 * linear loop. Locate the first node where r <= 1.
1076 */
1085 }
1092 }
1094 }
1098 }
1100 /*
1101 * These cases occur when the parent's idea of the boundary
1102 * is wider then the child's idea of the boundary, and
1103 * require special handling. If not inserting we can
1104 * terminate the search early for these cases but the
1105 * child's boundaries cannot be unconditionally modified.
1106 */
1108 /*
1109 * If i == 0 the search terminated to the LEFT of the
1110 * left_boundary but to the RIGHT of the parent's left
1111 * boundary.
1112 */
1113 u_int8_t save;
1117 /*
1118 * If we aren't inserting we can stop here.
1119 */
1124 }
1126 /*
1127 * Correct a left-hand boundary mismatch.
1128 *
1129 * We can only do this if we can upgrade the lock,
1130 * and synchronized as a background cursor (i.e.
1131 * inserting or pruning).
1132 *
1133 * WARNING: We can only do this if inserting, i.e.
1134 * we are running on the backend.
1135 */
1146 /*
1147 * If i == node->count + 1 the search terminated to
1148 * the RIGHT of the right boundary but to the LEFT
1149 * of the parent's right boundary. If we aren't
1150 * inserting we can stop here.
1151 *
1152 * Note that the last element in this case is
1153 * elms[i-2] prior to adjustments to 'i'.
1154 */
1160 }
1162 /*
1163 * Correct a right-hand boundary mismatch.
1164 * (actual push-down record is i-2 prior to
1165 * adjustments to i).
1166 *
1167 * We can only do this if we can upgrade the lock,
1168 * and synchronized as a background cursor (i.e.
1169 * inserting or pruning).
1170 *
1171 * WARNING: We can only do this if inserting, i.e.
1172 * we are running on the backend.
1173 */
1184 /*
1185 * The push-down index is now i - 1. If we had
1186 * terminated on the right boundary this will point
1187 * us at the last element.
1188 */
1190 }
1198 i,
1204 );
1205 }
1207 /*
1208 * We better have a valid subtree offset.
1209 */
1212 /*
1213 * Handle insertion and deletion requirements.
1214 *
1215 * If inserting split full nodes. The split code will
1216 * adjust cursor->node and cursor->index if the current
1217 * index winds up in the new node.
1218 *
1219 * If inserting and a left or right edge case was detected,
1220 * we cannot correct the left or right boundary and must
1221 * prepend and append an empty leaf node in order to make
1222 * the boundary correction.
1223 *
1224 * If we run out of space we set enospc and continue on
1225 * to a leaf to provide the spike code with a good point
1226 * of entry.
1227 */
1235 }
1236 /*
1237 * reload stale pointers
1238 */
1241 }
1242 }
1244 /*
1245 * Push down (push into new node, existing node becomes
1246 * the parent) and continue the search.
1247 */
1249 /* node may have become stale */
1253 }
1255 /*
1256 * We are at a leaf, do a linear search of the key array.
1257 *
1258 * On success the index is set to the matching element and 0
1259 * is returned.
1260 *
1261 * On failure the index is set to the insertion point and ENOENT
1262 * is returned.
1263 *
1264 * Boundaries are not stored in leaf nodes, so the index can wind
1265 * up to the left of element 0 (index == 0) or past the end of
1266 * the array (index == node->count). It is also possible that the
1267 * leaf might be empty.
1268 */
1276 }
1278 /*
1279 * Try to shortcut the search before dropping into the
1280 * linear loop. Locate the first node where r <= 1.
1281 */
1292 /*
1293 * We are at a record element. Stop if we've flipped past
1294 * key_beg, not counting the create_tid test. Allow the
1295 * r == 1 case (key_beg > element but differs only by its
1296 * create_tid) to fall through to the AS-OF check.
1297 */
1305 }
1307 /*
1308 * Check our as-of timestamp against the element.
1309 */
1315 }
1316 /* success */
1321 }
1322 /* success */
1323 }
1329 }
1331 }
1333 /*
1334 * The search of the leaf node failed. i is the insertion point.
1335 */
1336 failed:
1340 }
1342 /*
1343 * No exact match was found, i is now at the insertion point.
1344 *
1345 * If inserting split a full leaf before returning. This
1346 * may have the side effect of adjusting cursor->node and
1347 * cursor->index.
1348 */
1357 }
1358 /*
1359 * reload stale pointers
1360 */
1361 /* NOT USED
1362 i = cursor->index;
1363 node = &cursor->node->internal;
1364 */
1365 }
1367 /*
1368 * We reached a leaf but did not find the key we were looking for.
1369 * If this is an insert we will be properly positioned for an insert
1370 * (ENOENT) or spike (ENOSPC) operation.
1371 */
1373 done:
1375 }
1377 /*
1378 * Heuristical search for the first element whos comparison is <= 1. May
1379 * return an index whos compare result is > 1 but may only return an index
1380 * whos compare result is <= 1 if it is the first element with that result.
1381 */
1382 int
1384 {
1390 /*
1391 * Don't bother if the node does not have very many elements
1392 */
1403 }
1404 }
1406 }
1409 /************************************************************************
1410 * SPLITTING AND MERGING *
1411 ************************************************************************
1412 *
1413 * These routines do all the dirty work required to split and merge nodes.
1414 */
1416 /*
1417 * Split an internal node into two nodes and move the separator at the split
1418 * point to the parent.
1419 *
1420 * (cursor->node, cursor->index) indicates the element the caller intends
1421 * to push into. We will adjust node and index if that element winds
1422 * up in the split node.
1423 *
1424 * If we are at the root of the filesystem a new root must be created with
1425 * two elements, one pointing to the original root and one pointing to the
1426 * newly allocated split node.
1427 */
1428 static
1429 int
1431 {
1432 hammer_node_ondisk_t ondisk;
1433 hammer_node_t node;
1434 hammer_node_t parent;
1435 hammer_node_t new_node;
1436 hammer_btree_elm_t elm;
1437 hammer_btree_elm_t parent_elm;
1440 hammer_off_t hint;
1456 /*
1457 * We are splitting but elms[split] will be promoted to the parent,
1458 * leaving the right hand node with one less element. If the
1459 * insertion point will be on the left-hand side adjust the split
1460 * point to give the right hand side one additional node.
1461 */
1468 /*
1469 * If we are at the root of the filesystem, create a new root node
1470 * with 1 element and split normally. Avoid making major
1471 * modifications until we know the whole operation will work.
1472 */
1490 /* ondisk->elms[1].base.btype - not used */
1497 }
1499 /*
1500 * Calculate a hint for the allocation of the new B-Tree node.
1501 * The most likely expansion is coming from the insertion point
1502 * at cursor->index, so try to localize the allocation of our
1503 * new node to accomodate that sub-tree.
1504 *
1505 * Use the right-most sub-tree when expandinging on the right edge.
1506 * This is a very common case when copying a directory tree.
1507 */
1510 else
1513 /*
1514 * Split node into new_node at the split point.
1515 *
1516 * B O O O P N N B <-- P = node->elms[split] (index 4)
1517 * 0 1 2 3 4 5 6 <-- subtree indices
1518 *
1519 * x x P x x
1520 * s S S s
1521 * / \
1522 * B O O O B B N N B <--- inner boundary points are 'P'
1523 * 0 1 2 3 4 5 6
1524 */
1531 }
1533 }
1536 /*
1537 * Create the new node. P becomes the left-hand boundary in the
1538 * new node. Copy the right-hand boundary as well.
1539 *
1540 * elm is the new separator.
1541 */
1555 /*
1556 * Cleanup the original node. Elm (P) becomes the new boundary,
1557 * its subtree_offset was moved to the new node. If we had created
1558 * a new root its parent pointer may have changed.
1559 */
1563 /*
1564 * Insert the separator into the parent, fixup the parent's
1565 * reference to the original node, and reference the new node.
1566 * The separator is P.
1567 *
1568 * Remember that base.count does not include the right-hand boundary.
1569 */
1584 /*
1585 * The children of new_node need their parent pointer set to new_node.
1586 * The children have already been locked by
1587 * hammer_btree_lock_children().
1588 */
1594 }
1595 }
1598 /*
1599 * The filesystem's root B-Tree pointer may have to be updated.
1600 */
1602 hammer_volume_t volume;
1608 vol0_btree_root);
1615 }
1618 }
1621 /*
1622 * Ok, now adjust the cursor depending on which element the original
1623 * index was pointing at. If we are >= the split point the push node
1624 * is now in the new node.
1625 *
1626 * NOTE: If we are at the split point itself we cannot stay with the
1627 * original node because the push index will point at the right-hand
1628 * boundary, which is illegal.
1629 *
1630 * NOTE: The cursor's parent or parent_index must be adjusted for
1631 * the case where a new parent (new root) was created, and the case
1632 * where the cursor is now pointing at the split node.
1633 */
1644 }
1646 /*
1647 * Fixup left and right bounds
1648 */
1657 done:
1661 }
1663 /*
1664 * Same as the above, but splits a full leaf node.
1665 *
1666 * This function
1667 */
1668 static
1669 int
1671 {
1672 hammer_node_ondisk_t ondisk;
1673 hammer_node_t parent;
1674 hammer_node_t leaf;
1675 hammer_mount_t hmp;
1676 hammer_node_t new_leaf;
1677 hammer_btree_elm_t elm;
1678 hammer_btree_elm_t parent_elm;
1679 hammer_base_elm_t mid_boundary;
1680 hammer_off_t hint;
1696 /*
1697 * Calculate the split point. If the insertion point will be on
1698 * the left-hand side adjust the split point to give the right
1699 * hand side one additional node.
1700 *
1701 * Spikes are made up of two leaf elements which cannot be
1702 * safely split.
1703 */
1719 /*
1720 * If we are at the root of the tree, create a new root node with
1721 * 1 element and split normally. Avoid making major modifications
1722 * until we know the whole operation will work.
1723 */
1740 /* ondisk->elms[1].base.btype = not used */
1748 }
1750 /*
1751 * Calculate a hint for the allocation of the new B-Tree leaf node.
1752 * For now just try to localize it within the same bigblock as
1753 * the current leaf.
1754 *
1755 * If the insertion point is at the end of the leaf we recognize a
1756 * likely append sequence of some sort (data, meta-data, inodes,
1757 * whatever). Set the hint to zero to allocate out of linear space
1758 * instead of trying to completely fill previously hinted space.
1759 *
1760 * This also sets the stage for recursive splits to localize using
1761 * the new space.
1762 */
1766 else
1769 /*
1770 * Split leaf into new_leaf at the split point. Select a separator
1771 * value in-between the two leafs but with a bent towards the right
1772 * leaf since comparisons use an 'elm >= separator' inequality.
1773 *
1774 * L L L L L L L L
1775 *
1776 * x x P x x
1777 * s S S s
1778 * / \
1779 * L L L L L L L L
1780 */
1787 }
1789 }
1792 /*
1793 * Create the new node and copy the leaf elements from the split
1794 * point on to the new node.
1795 */
1809 /*
1810 * Cleanup the original node. Because this is a leaf node and
1811 * leaf nodes do not have a right-hand boundary, there
1812 * aren't any special edge cases to clean up. We just fixup the
1813 * count.
1814 */
1817 /*
1818 * Insert the separator into the parent, fixup the parent's
1819 * reference to the original node, and reference the new node.
1820 * The separator is P.
1821 *
1822 * Remember that base.count does not include the right-hand boundary.
1823 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1824 */
1842 /*
1843 * The filesystem's root B-Tree pointer may have to be updated.
1844 */
1846 hammer_volume_t volume;
1852 vol0_btree_root);
1859 }
1862 }
1865 /*
1866 * Ok, now adjust the cursor depending on which element the original
1867 * index was pointing at. If we are >= the split point the push node
1868 * is now in the new node.
1869 *
1870 * NOTE: If we are at the split point itself we need to select the
1871 * old or new node based on where key_beg's insertion point will be.
1872 * If we pick the wrong side the inserted element will wind up in
1873 * the wrong leaf node and outside that node's bounds.
1874 */
1887 }
1889 /*
1890 * Fixup left and right bounds
1891 */
1896 /*
1897 * Assert that the bounds are correct.
1898 */
1906 done:
1909 }
1911 #if 0
1913 /*
1914 * Recursively correct the right-hand boundary's create_tid to (tid) as
1915 * long as the rest of the key matches. We have to recurse upward in
1916 * the tree as well as down the left side of each parent's right node.
1917 *
1918 * Return EDEADLK if we were only partially successful, forcing the caller
1919 * to try again. The original cursor is not modified. This routine can
1920 * also fail with EDEADLK if it is forced to throw away a portion of its
1921 * record history.
1922 *
1923 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1924 */
1927 hammer_node_t node;
1929 };
1933 int
1935 {
1938 hammer_base_elm_t elm;
1939 hammer_node_t orig_node;
1947 /*
1948 * Save our position so we can restore it on return. This also
1949 * gives us a stable 'elm'.
1950 */
1957 /*
1958 * Now build a list of parents going up, allocating a rhb
1959 * structure for each one.
1960 */
1962 /*
1963 * Stop if we no longer have any right-bounds to fix up
1964 */
1969 }
1971 /*
1972 * Stop if the right-hand bound's create_tid does not
1973 * need to be corrected.
1974 */
1986 }
1988 /*
1989 * now safely adjust the right hand bound for each rhb. This may
1990 * also require taking the right side of the tree and iterating down
1991 * ITS left side.
1992 */
2005 /*
2006 * Right-boundary for parent at internal node
2007 * is one element to the right of the element whos
2008 * right boundary needs adjusting. We must then
2009 * traverse down the left side correcting any left
2010 * bounds (which may now be too far to the left).
2011 */
2019 }
2020 }
2022 /*
2023 * Cleanup
2024 */
2030 }
2035 }
2037 /*
2038 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2039 * bound going downward starting at the current cursor position.
2040 *
2041 * This function does not restore the cursor after use.
2042 */
2043 int
2045 {
2047 hammer_base_elm_t elm;
2048 hammer_base_elm_t cmp;
2058 /*
2059 * Record the node and traverse down the left-hand side for all
2060 * matching records needing a boundary correction.
2061 */
2072 /*
2073 * Nothing to traverse down if we are at the right
2074 * boundary of an internal node.
2075 */
2083 }
2093 }
2097 }
2099 /*
2100 * Now we can safely adjust the left-hand boundary from the bottom-up.
2101 * The last element we remove from the list is the caller's right hand
2102 * boundary, which must also be adjusted.
2103 */
2122 }
2123 }
2125 /*
2126 * Cleanup
2127 */
2133 }
2135 }
2137 #endif
2139 /*
2140 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2141 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2142 * the operation due to a deadlock, or some other error.
2143 *
2144 * This routine is initially called with an empty leaf and may be
2145 * recursively called with single-element internal nodes.
2146 *
2147 * It should also be noted that when removing empty leaves we must be sure
2148 * to test and update mirror_tid because another thread may have deadlocked
2149 * against us (or someone) trying to propagate it up and cannot retry once
2150 * the node has been deleted.
2151 *
2152 * On return the cursor may end up pointing to an internal node, suitable
2153 * for further iteration but not for an immediate insertion or deletion.
2154 */
2155 static int
2157 {
2158 hammer_node_ondisk_t ondisk;
2159 hammer_btree_elm_t elm;
2160 hammer_node_t node;
2161 hammer_node_t parent;
2167 /*
2168 * When deleting the root of the filesystem convert it to
2169 * an empty leaf node. Internal nodes cannot be empty.
2170 */
2181 }
2186 /*
2187 * Attempt to remove the parent's reference to the child. If the
2188 * parent would become empty we have to recurse. If we fail we
2189 * leave the parent pointing to an empty leaf node.
2190 *
2191 * We have to recurse successfully before we can delete the internal
2192 * node as it is illegal to have empty internal nodes. Even though
2193 * the operation may be aborted we must still fixup any unlocked
2194 * cursors as if we had deleted the element prior to recursing
2195 * (by calling hammer_cursor_deleted_element()) so those cursors
2196 * are properly forced up the chain by the recursion.
2197 */
2199 /*
2200 * This special cursor_up_locked() call leaves the original
2201 * node exclusively locked and referenced, leaves the
2202 * original parent locked (as the new node), and locks the
2203 * new parent. It can return EDEADLK.
2204 */
2218 /*
2219 * Defer parent removal because we could not
2220 * get the lock, just let the leaf remain
2221 * empty.
2222 */
2223 /**/
2224 }
2228 /*
2229 * Defer parent removal because we could not
2230 * get the lock, just let the leaf remain
2231 * empty.
2232 */
2233 /**/
2234 }
2246 /*
2247 * We must retain the highest mirror_tid. The deleted
2248 * range is now encompassed by the element to the left.
2249 * If we are already at the left edge the new left edge
2250 * inherits mirror_tid.
2251 *
2252 * Note that bounds of the parent to our parent may create
2253 * a gap to the left of our left-most node or to the right
2254 * of our right-most node. The gap is silently included
2255 * in the mirror_tid's area of effect from the point of view
2256 * of the scan.
2257 */
2263 }
2269 }
2270 }
2272 /*
2273 * Delete the subtree reference in the parent
2274 */
2283 /*
2284 * cursor->node is invalid, cursor up to make the cursor
2285 * valid again.
2286 */
2288 }
2290 }
2292 /*
2293 * Propagate cursor->trans->tid up the B-Tree starting at the current
2294 * cursor position using pseudofs info gleaned from the passed inode.
2295 *
2296 * The passed inode has no relationship to the cursor position other
2297 * then being in the same pseudofs as the insertion or deletion we
2298 * are propagating the mirror_tid for.
2299 */
2300 void
2302 hammer_pseudofs_inmem_t pfsm,
2303 hammer_btree_leaf_elm_t leaf)
2304 {
2305 hammer_cursor_t ncursor;
2306 hammer_tid_t mirror_tid;
2309 /*
2310 * We do not propagate a mirror_tid if the filesystem was mounted
2311 * in no-mirror mode.
2312 */
2316 /*
2317 * This is a bit of a hack because we cannot deadlock or return
2318 * EDEADLK here. The related operation has already completed and
2319 * we must propagate the mirror_tid now regardless.
2320 *
2321 * Generate a new cursor which inherits the original's locks and
2322 * unlock the original. Use the new cursor to propagate the
2323 * mirror_tid. Then clean up the new cursor and reacquire locks
2324 * on the original.
2325 *
2326 * hammer_dup_cursor() cannot dup locks. The dup inherits the
2327 * original's locks and the original is tracked and must be
2328 * re-locked.
2329 */
2336 }
2339 /*
2340 * Propagate a mirror TID update upwards through the B-Tree to the root.
2341 *
2342 * A locked internal node must be passed in. The node will remain locked
2343 * on return.
2344 *
2345 * This function syncs mirror_tid at the specified internal node's element,
2346 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2347 */
2348 static int
2350 {
2351 hammer_btree_internal_elm_t elm;
2352 hammer_node_t node;
2362 }
2368 /*
2369 * Adjust the node's element
2370 */
2384 }
2387 /*
2388 * Adjust the node's mirror_tid aggregator
2389 */
2400 }
2401 }
2405 }
2407 hammer_node_t
2410 {
2411 hammer_node_t parent;
2412 hammer_btree_elm_t elm;
2415 /*
2416 * Get the node
2417 */
2422 }
2425 /*
2426 * Lock the node
2427 */
2433 }
2436 }
2438 /*
2439 * Figure out which element in the parent is pointing to the
2440 * child.
2441 */
2447 }
2453 }
2457 }
2461 }
2463 /*
2464 * The element (elm) has been moved to a new internal node (node).
2465 *
2466 * If the element represents a pointer to an internal node that node's
2467 * parent must be adjusted to the element's new location.
2468 *
2469 * XXX deadlock potential here with our exclusive locks
2470 */
2471 int
2473 hammer_btree_elm_t elm)
2474 {
2475 hammer_node_t child;
2490 }
2494 }
2496 }
2498 /*
2499 * Initialize the root of a recursive B-Tree node lock list structure.
2500 */
2501 void
2503 {
2511 }
2513 /*
2514 * Exclusively lock all the children of node. This is used by the split
2515 * code to prevent anyone from accessing the children of a cursor node
2516 * while we fix-up its parent offset.
2517 *
2518 * If we don't lock the children we can really mess up cursors which block
2519 * trying to cursor-up into our node.
2520 *
2521 * On failure EDEADLK (or some other error) is returned. If a deadlock
2522 * error is returned the cursor is adjusted to block on termination.
2523 *
2524 * The caller is responsible for managing parent->node, the root's node
2525 * is usually aliased from a cursor.
2526 */
2527 int
2529 hammer_node_lock_t parent)
2530 {
2531 hammer_node_t node;
2532 hammer_node_lock_t item;
2533 hammer_node_ondisk_t ondisk;
2534 hammer_btree_elm_t elm;
2535 hammer_node_t child;
2545 /*
2546 * We really do not want to block on I/O with exclusive locks held,
2547 * pre-get the children before trying to lock the mess. This is
2548 * only done one-level deep for now.
2549 */
2556 }
2562 }
2564 /*
2565 * Do it for real
2566 */
2582 }
2588 }
2601 /*
2602 * Recurse (used by the rebalancing code)
2603 */
2606 cursor,
2608 item);
2609 }
2610 }
2611 }
2612 }
2616 }
2618 /*
2619 * Create an in-memory copy of all B-Tree nodes listed, recursively,
2620 * including the parent.
2621 */
2622 void
2624 {
2626 hammer_node_lock_t item;
2630 M_WAITOK);
2632 }
2635 }
2636 }
2638 /*
2639 * Recursively sync modified copies to the media.
2640 */
2641 int
2643 {
2644 hammer_node_lock_t item;
2655 }
2656 }
2659 }
2661 }
2663 /*
2664 * Release previously obtained node locks. The caller is responsible for
2665 * cleaning up parent->node itself (its usually just aliased from a cursor),
2666 * but this function will take care of the copies.
2667 */
2668 void
2670 {
2671 hammer_node_lock_t item;
2676 }
2683 }
2684 }
2686 /************************************************************************
2687 * MISCELLANIOUS SUPPORT *
2688 ************************************************************************/
2690 /*
2691 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2692 *
2693 * Note that for this particular function a return value of -1, 0, or +1
2694 * can denote a match if create_tid is otherwise discounted. A create_tid
2695 * of zero is considered to be 'infinity' in comparisons.
2696 *
2697 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2698 */
2699 int
2701 {
2722 /*
2723 * A create_tid of zero indicates a record which is undeletable
2724 * and must be considered to have a value of positive infinity.
2725 */
2730 }
2738 }
2740 /*
2741 * Test a timestamp against an element to determine whether the
2742 * element is visible. A timestamp of 0 means 'infinity'.
2743 */
2744 int
2746 {
2751 }
2757 }
2759 /*
2760 * Create a separator half way inbetween key1 and key2. For fields just
2761 * one unit apart, the separator will match key2. key1 is on the left-hand
2762 * side and key2 is on the right-hand side.
2763 *
2764 * key2 must be >= the separator. It is ok for the separator to match key2.
2765 *
2766 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2767 * key2.
2768 *
2769 * NOTE: It might be beneficial to just scrap this whole mess and just
2770 * set the separator to key2.
2771 */
2772 #define MAKE_SEPARATOR(key1, key2, dest, field) \
2773 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2775 static void
2777 hammer_base_elm_t dest)
2778 {
2793 /*
2794 * Don't bother creating a separator for
2795 * create_tid, which also conveniently avoids
2796 * having to handle the create_tid == 0
2797 * (infinity) case. Just leave create_tid
2798 * set to key2.
2799 *
2800 * Worst case, dest matches key2 exactly,
2801 * which is acceptable.
2802 */
2803 }
2804 }
2805 }
2806 }
2808 #undef MAKE_SEPARATOR
2810 /*
2811 * Return whether a generic internal or leaf node is full
2812 */
2813 static int
2815 {
2827 }
2829 }
2831 #if 0
2832 static int
2834 {
2840 }
2841 #endif
2843 void
2845 {
2846 hammer_btree_elm_t elm;
2853 /*
2854 * Dump both boundary elements if an internal node
2855 */
2860 }
2865 }
2866 }
2867 }
2869 void
2871 {
2895 }
2896 }