HAMMER VFS - Implement swapcache for HAMMER data in double_buffer mode
[dragonfly.git] / sys / vfs / hammer / hammer_btree.c
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  */
36
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  */
81 #include "hammer.h"
82 #include <sys/buf.h>
83 #include <sys/buf2.h>
84
85 static int btree_search(hammer_cursor_t cursor, int flags);
86 static int btree_split_internal(hammer_cursor_t cursor);
87 static int btree_split_leaf(hammer_cursor_t cursor);
88 static int btree_remove(hammer_cursor_t cursor);
89 static int btree_node_is_full(hammer_node_ondisk_t node);
90 static int hammer_btree_mirror_propagate(hammer_cursor_t cursor,        
91                         hammer_tid_t mirror_tid);
92 static void hammer_make_separator(hammer_base_elm_t key1,
93                         hammer_base_elm_t key2, hammer_base_elm_t dest);
94 static void hammer_cursor_mirror_filter(hammer_cursor_t cursor);
95
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  * If HAMMER_CURSOR_ITERATE_CHECK is set it is possible that the cursor
114  * was reverse indexed due to being moved to a parent while unlocked,
115  * and something else might have inserted an element outside the iteration
116  * range.  When this case occurs the iterator just keeps iterating until
117  * it gets back into the iteration range (instead of asserting).
118  *
119  * NOTE!  EDEADLK *CANNOT* be returned by this procedure.
120  */
121 int
122 hammer_btree_iterate(hammer_cursor_t cursor)
123 {
124         hammer_node_ondisk_t node;
125         hammer_btree_elm_t elm;
126         hammer_mount_t hmp;
127         int error = 0;
128         int r;
129         int s;
130
131         /*
132          * Skip past the current record
133          */
134         hmp = cursor->trans->hmp;
135         node = cursor->node->ondisk;
136         if (node == NULL)
137                 return(ENOENT);
138         if (cursor->index < node->count && 
139             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
140                 ++cursor->index;
141         }
142
143         /*
144          * HAMMER can wind up being cpu-bound.
145          */
146         if (++hmp->check_yield > hammer_yield_check) {
147                 hmp->check_yield = 0;
148                 lwkt_user_yield();
149         }
150
151
152         /*
153          * Loop until an element is found or we are done.
154          */
155         for (;;) {
156                 /*
157                  * We iterate up the tree and then index over one element
158                  * while we are at the last element in the current node.
159                  *
160                  * If we are at the root of the filesystem, cursor_up
161                  * returns ENOENT.
162                  *
163                  * XXX this could be optimized by storing the information in
164                  * the parent reference.
165                  *
166                  * XXX we can lose the node lock temporarily, this could mess
167                  * up our scan.
168                  */
169                 ++hammer_stats_btree_iterations;
170                 hammer_flusher_clean_loose_ios(hmp);
171
172                 if (cursor->index == node->count) {
173                         if (hammer_debug_btree) {
174                                 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
175                                         (long long)cursor->node->node_offset,
176                                         cursor->index,
177                                         (long long)(cursor->parent ? cursor->parent->node_offset : -1),
178                                         cursor->parent_index,
179                                         curthread);
180                         }
181                         KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
182                         error = hammer_cursor_up(cursor);
183                         if (error)
184                                 break;
185                         /* reload stale pointer */
186                         node = cursor->node->ondisk;
187                         KKASSERT(cursor->index != node->count);
188
189                         /*
190                          * If we are reblocking we want to return internal
191                          * nodes.  Note that the internal node will be
192                          * returned multiple times, on each upward recursion
193                          * from its children.  The caller selects which
194                          * revisit it cares about (usually first or last only).
195                          */
196                         if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
197                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
198                                 return(0);
199                         }
200                         ++cursor->index;
201                         continue;
202                 }
203
204                 /*
205                  * Check internal or leaf element.  Determine if the record
206                  * at the cursor has gone beyond the end of our range.
207                  *
208                  * We recurse down through internal nodes.
209                  */
210                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
211                         elm = &node->elms[cursor->index];
212
213                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
214                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
215                         if (hammer_debug_btree) {
216                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
217                                         (long long)cursor->node->node_offset,
218                                         cursor->index,
219                                         (long long)elm[0].internal.base.obj_id,
220                                         elm[0].internal.base.rec_type,
221                                         (long long)elm[0].internal.base.key,
222                                         elm[0].internal.base.localization,
223                                         r,
224                                         curthread
225                                 );
226                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
227                                         (long long)cursor->node->node_offset,
228                                         cursor->index + 1,
229                                         (long long)elm[1].internal.base.obj_id,
230                                         elm[1].internal.base.rec_type,
231                                         (long long)elm[1].internal.base.key,
232                                         elm[1].internal.base.localization,
233                                         s
234                                 );
235                         }
236
237                         if (r < 0) {
238                                 error = ENOENT;
239                                 break;
240                         }
241                         if (r == 0 && (cursor->flags &
242                                        HAMMER_CURSOR_END_INCLUSIVE) == 0) {
243                                 error = ENOENT;
244                                 break;
245                         }
246
247                         /*
248                          * Better not be zero
249                          */
250                         KKASSERT(elm->internal.subtree_offset != 0);
251
252                         if (s <= 0) {
253                                 /*
254                                  * If running the mirror filter see if we
255                                  * can skip one or more entire sub-trees.
256                                  * If we can we return the internal node
257                                  * and the caller processes the skipped
258                                  * range (see mirror_read).
259                                  */
260                                 if (cursor->flags &
261                                     HAMMER_CURSOR_MIRROR_FILTERED) {
262                                         if (elm->internal.mirror_tid <
263                                             cursor->cmirror->mirror_tid) {
264                                                 hammer_cursor_mirror_filter(cursor);
265                                                 return(0);
266                                         }
267                                 }
268                         } else {
269                                 /*
270                                  * Normally it would be impossible for the
271                                  * cursor to have gotten back-indexed,
272                                  * but it can happen if a node is deleted
273                                  * and the cursor is moved to its parent
274                                  * internal node.  ITERATE_CHECK will be set.
275                                  */
276                                 KKASSERT(cursor->flags &
277                                          HAMMER_CURSOR_ITERATE_CHECK);
278                                 kprintf("hammer_btree_iterate: "
279                                         "DEBUG: Caught parent seek "
280                                         "in internal iteration\n");
281                         }
282
283                         error = hammer_cursor_down(cursor);
284                         if (error)
285                                 break;
286                         KKASSERT(cursor->index == 0);
287                         /* reload stale pointer */
288                         node = cursor->node->ondisk;
289                         continue;
290                 } else {
291                         elm = &node->elms[cursor->index];
292                         r = hammer_btree_cmp(&cursor->key_end, &elm->base);
293                         if (hammer_debug_btree) {
294                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
295                                         (long long)cursor->node->node_offset,
296                                         cursor->index,
297                                         (elm[0].leaf.base.btype ?
298                                          elm[0].leaf.base.btype : '?'),
299                                         (long long)elm[0].leaf.base.obj_id,
300                                         elm[0].leaf.base.rec_type,
301                                         (long long)elm[0].leaf.base.key,
302                                         elm[0].leaf.base.localization,
303                                         r
304                                 );
305                         }
306                         if (r < 0) {
307                                 error = ENOENT;
308                                 break;
309                         }
310
311                         /*
312                          * We support both end-inclusive and
313                          * end-exclusive searches.
314                          */
315                         if (r == 0 &&
316                            (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
317                                 error = ENOENT;
318                                 break;
319                         }
320
321                         /*
322                          * If ITERATE_CHECK is set an unlocked cursor may
323                          * have been moved to a parent and the iterate can
324                          * happen upon elements that are not in the requested
325                          * range.
326                          */
327                         if (cursor->flags & HAMMER_CURSOR_ITERATE_CHECK) {
328                                 s = hammer_btree_cmp(&cursor->key_beg,
329                                                      &elm->base);
330                                 if (s > 0) {
331                                         kprintf("hammer_btree_iterate: "
332                                                 "DEBUG: Caught parent seek "
333                                                 "in leaf iteration\n");
334                                         ++cursor->index;
335                                         continue;
336                                 }
337                         }
338                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
339
340                         /*
341                          * Return the element
342                          */
343                         switch(elm->leaf.base.btype) {
344                         case HAMMER_BTREE_TYPE_RECORD:
345                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
346                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
347                                         ++cursor->index;
348                                         continue;
349                                 }
350                                 error = 0;
351                                 break;
352                         default:
353                                 error = EINVAL;
354                                 break;
355                         }
356                         if (error)
357                                 break;
358                 }
359                 /*
360                  * node pointer invalid after loop
361                  */
362
363                 /*
364                  * Return entry
365                  */
366                 if (hammer_debug_btree) {
367                         int i = cursor->index;
368                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
369                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
370                                 cursor->node, i,
371                                 (long long)elm->internal.base.obj_id,
372                                 elm->internal.base.rec_type,
373                                 (long long)elm->internal.base.key,
374                                 elm->internal.base.localization
375                         );
376                 }
377                 return(0);
378         }
379         return(error);
380 }
381
382 /*
383  * We hit an internal element that we could skip as part of a mirroring
384  * scan.  Calculate the entire range being skipped.
385  *
386  * It is important to include any gaps between the parent's left_bound
387  * and the node's left_bound, and same goes for the right side.
388  */
389 static void
390 hammer_cursor_mirror_filter(hammer_cursor_t cursor)
391 {
392         struct hammer_cmirror *cmirror;
393         hammer_node_ondisk_t ondisk;
394         hammer_btree_elm_t elm;
395
396         ondisk = cursor->node->ondisk;
397         cmirror = cursor->cmirror;
398
399         /*
400          * Calculate the skipped range
401          */
402         elm = &ondisk->elms[cursor->index];
403         if (cursor->index == 0)
404                 cmirror->skip_beg = *cursor->left_bound;
405         else
406                 cmirror->skip_beg = elm->internal.base;
407         while (cursor->index < ondisk->count) {
408                 if (elm->internal.mirror_tid >= cmirror->mirror_tid)
409                         break;
410                 ++cursor->index;
411                 ++elm;
412         }
413         if (cursor->index == ondisk->count)
414                 cmirror->skip_end = *cursor->right_bound;
415         else
416                 cmirror->skip_end = elm->internal.base;
417
418         /*
419          * clip the returned result.
420          */
421         if (hammer_btree_cmp(&cmirror->skip_beg, &cursor->key_beg) < 0)
422                 cmirror->skip_beg = cursor->key_beg;
423         if (hammer_btree_cmp(&cmirror->skip_end, &cursor->key_end) > 0)
424                 cmirror->skip_end = cursor->key_end;
425 }
426
427 /*
428  * Iterate in the reverse direction.  This is used by the pruning code to
429  * avoid overlapping records.
430  */
431 int
432 hammer_btree_iterate_reverse(hammer_cursor_t cursor)
433 {
434         hammer_node_ondisk_t node;
435         hammer_btree_elm_t elm;
436         hammer_mount_t hmp;
437         int error = 0;
438         int r;
439         int s;
440
441         /* mirror filtering not supported for reverse iteration */
442         KKASSERT ((cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) == 0);
443
444         /*
445          * Skip past the current record.  For various reasons the cursor
446          * may end up set to -1 or set to point at the end of the current
447          * node.  These cases must be addressed.
448          */
449         node = cursor->node->ondisk;
450         if (node == NULL)
451                 return(ENOENT);
452         if (cursor->index != -1 && 
453             (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
454                 --cursor->index;
455         }
456         if (cursor->index == cursor->node->ondisk->count)
457                 --cursor->index;
458
459         /*
460          * HAMMER can wind up being cpu-bound.
461          */
462         hmp = cursor->trans->hmp;
463         if (++hmp->check_yield > hammer_yield_check) {
464                 hmp->check_yield = 0;
465                 lwkt_user_yield();
466         }
467
468         /*
469          * Loop until an element is found or we are done.
470          */
471         for (;;) {
472                 ++hammer_stats_btree_iterations;
473                 hammer_flusher_clean_loose_ios(hmp);
474
475                 /*
476                  * We iterate up the tree and then index over one element
477                  * while we are at the last element in the current node.
478                  */
479                 if (cursor->index == -1) {
480                         error = hammer_cursor_up(cursor);
481                         if (error) {
482                                 cursor->index = 0; /* sanity */
483                                 break;
484                         }
485                         /* reload stale pointer */
486                         node = cursor->node->ondisk;
487                         KKASSERT(cursor->index != node->count);
488                         --cursor->index;
489                         continue;
490                 }
491
492                 /*
493                  * Check internal or leaf element.  Determine if the record
494                  * at the cursor has gone beyond the end of our range.
495                  *
496                  * We recurse down through internal nodes. 
497                  */
498                 KKASSERT(cursor->index != node->count);
499                 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
500                         elm = &node->elms[cursor->index];
501                         r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
502                         s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
503                         if (hammer_debug_btree) {
504                                 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
505                                         (long long)cursor->node->node_offset,
506                                         cursor->index,
507                                         (long long)elm[0].internal.base.obj_id,
508                                         elm[0].internal.base.rec_type,
509                                         (long long)elm[0].internal.base.key,
510                                         elm[0].internal.base.localization,
511                                         r
512                                 );
513                                 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
514                                         (long long)cursor->node->node_offset,
515                                         cursor->index + 1,
516                                         (long long)elm[1].internal.base.obj_id,
517                                         elm[1].internal.base.rec_type,
518                                         (long long)elm[1].internal.base.key,
519                                         elm[1].internal.base.localization,
520                                         s
521                                 );
522                         }
523
524                         if (s >= 0) {
525                                 error = ENOENT;
526                                 break;
527                         }
528
529                         /*
530                          * It shouldn't be possible to be seeked past key_end,
531                          * even if the cursor got moved to a parent.
532                          */
533                         KKASSERT(r >= 0);
534
535                         /*
536                          * Better not be zero
537                          */
538                         KKASSERT(elm->internal.subtree_offset != 0);
539
540                         error = hammer_cursor_down(cursor);
541                         if (error)
542                                 break;
543                         KKASSERT(cursor->index == 0);
544                         /* reload stale pointer */
545                         node = cursor->node->ondisk;
546
547                         /* this can assign -1 if the leaf was empty */
548                         cursor->index = node->count - 1;
549                         continue;
550                 } else {
551                         elm = &node->elms[cursor->index];
552                         s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
553                         if (hammer_debug_btree) {
554                                 kprintf("ELEMENT  %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
555                                         (long long)cursor->node->node_offset,
556                                         cursor->index,
557                                         (elm[0].leaf.base.btype ?
558                                          elm[0].leaf.base.btype : '?'),
559                                         (long long)elm[0].leaf.base.obj_id,
560                                         elm[0].leaf.base.rec_type,
561                                         (long long)elm[0].leaf.base.key,
562                                         elm[0].leaf.base.localization,
563                                         s
564                                 );
565                         }
566                         if (s > 0) {
567                                 error = ENOENT;
568                                 break;
569                         }
570
571                         /*
572                          * It shouldn't be possible to be seeked past key_end,
573                          * even if the cursor got moved to a parent.
574                          */
575                         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
576
577                         /*
578                          * Return the element
579                          */
580                         switch(elm->leaf.base.btype) {
581                         case HAMMER_BTREE_TYPE_RECORD:
582                                 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
583                                     hammer_btree_chkts(cursor->asof, &elm->base)) {
584                                         --cursor->index;
585                                         continue;
586                                 }
587                                 error = 0;
588                                 break;
589                         default:
590                                 error = EINVAL;
591                                 break;
592                         }
593                         if (error)
594                                 break;
595                 }
596                 /*
597                  * node pointer invalid after loop
598                  */
599
600                 /*
601                  * Return entry
602                  */
603                 if (hammer_debug_btree) {
604                         int i = cursor->index;
605                         hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
606                         kprintf("ITERATE  %p:%d %016llx %02x %016llx lo=%02x\n",
607                                 cursor->node, i,
608                                 (long long)elm->internal.base.obj_id,
609                                 elm->internal.base.rec_type,
610                                 (long long)elm->internal.base.key,
611                                 elm->internal.base.localization
612                         );
613                 }
614                 return(0);
615         }
616         return(error);
617 }
618
619 /*
620  * Lookup cursor->key_beg.  0 is returned on success, ENOENT if the entry
621  * could not be found, EDEADLK if inserting and a retry is needed, and a
622  * fatal error otherwise.  When retrying, the caller must terminate the
623  * cursor and reinitialize it.  EDEADLK cannot be returned if not inserting.
624  * 
625  * The cursor is suitably positioned for a deletion on success, and suitably
626  * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
627  * specified.
628  *
629  * The cursor may begin anywhere, the search will traverse the tree in
630  * either direction to locate the requested element.
631  *
632  * Most of the logic implementing historical searches is handled here.  We
633  * do an initial lookup with create_tid set to the asof TID.  Due to the
634  * way records are laid out, a backwards iteration may be required if
635  * ENOENT is returned to locate the historical record.  Here's the
636  * problem:
637  *
638  * create_tid:    10      15       20
639  *                   LEAF1   LEAF2
640  * records:         (11)        (18)
641  *
642  * Lets say we want to do a lookup AS-OF timestamp 17.  We will traverse
643  * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
644  * not visible and thus causes ENOENT to be returned.  We really need
645  * to check record 11 in LEAF1.  If it also fails then the search fails
646  * (e.g. it might represent the range 11-16 and thus still not match our
647  * AS-OF timestamp of 17).  Note that LEAF1 could be empty, requiring
648  * further iterations.
649  *
650  * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
651  * and the cursor->create_check TID if an iteration might be needed.
652  * In the above example create_check would be set to 14.
653  */
654 int
655 hammer_btree_lookup(hammer_cursor_t cursor)
656 {
657         int error;
658
659         cursor->flags &= ~HAMMER_CURSOR_ITERATE_CHECK;
660         KKASSERT ((cursor->flags & HAMMER_CURSOR_INSERT) == 0 ||
661                   cursor->trans->sync_lock_refs > 0);
662         ++hammer_stats_btree_lookups;
663         if (cursor->flags & HAMMER_CURSOR_ASOF) {
664                 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
665                 cursor->key_beg.create_tid = cursor->asof;
666                 for (;;) {
667                         cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
668                         error = btree_search(cursor, 0);
669                         if (error != ENOENT ||
670                             (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
671                                 /*
672                                  * Stop if no error.
673                                  * Stop if error other then ENOENT.
674                                  * Stop if ENOENT and not special case.
675                                  */
676                                 break;
677                         }
678                         if (hammer_debug_btree) {
679                                 kprintf("CREATE_CHECK %016llx\n",
680                                         (long long)cursor->create_check);
681                         }
682                         cursor->key_beg.create_tid = cursor->create_check;
683                         /* loop */
684                 }
685         } else {
686                 error = btree_search(cursor, 0);
687         }
688         if (error == 0)
689                 error = hammer_btree_extract(cursor, cursor->flags);
690         return(error);
691 }
692
693 /*
694  * Execute the logic required to start an iteration.  The first record
695  * located within the specified range is returned and iteration control
696  * flags are adjusted for successive hammer_btree_iterate() calls.
697  *
698  * Set ATEDISK so a low-level caller can call btree_first/btree_iterate
699  * in a loop without worrying about it.  Higher-level merged searches will
700  * adjust the flag appropriately.
701  */
702 int
703 hammer_btree_first(hammer_cursor_t cursor)
704 {
705         int error;
706
707         error = hammer_btree_lookup(cursor);
708         if (error == ENOENT) {
709                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
710                 error = hammer_btree_iterate(cursor);
711         }
712         cursor->flags |= HAMMER_CURSOR_ATEDISK;
713         return(error);
714 }
715
716 /*
717  * Similarly but for an iteration in the reverse direction.
718  *
719  * Set ATEDISK when iterating backwards to skip the current entry,
720  * which after an ENOENT lookup will be pointing beyond our end point.
721  *
722  * Set ATEDISK so a low-level caller can call btree_last/btree_iterate_reverse
723  * in a loop without worrying about it.  Higher-level merged searches will
724  * adjust the flag appropriately.
725  */
726 int
727 hammer_btree_last(hammer_cursor_t cursor)
728 {
729         struct hammer_base_elm save;
730         int error;
731
732         save = cursor->key_beg;
733         cursor->key_beg = cursor->key_end;
734         error = hammer_btree_lookup(cursor);
735         cursor->key_beg = save;
736         if (error == ENOENT ||
737             (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
738                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
739                 error = hammer_btree_iterate_reverse(cursor);
740         }
741         cursor->flags |= HAMMER_CURSOR_ATEDISK;
742         return(error);
743 }
744
745 /*
746  * Extract the record and/or data associated with the cursor's current
747  * position.  Any prior record or data stored in the cursor is replaced.
748  * The cursor must be positioned at a leaf node.
749  *
750  * NOTE: All extractions occur at the leaf of the B-Tree.
751  */
752 int
753 hammer_btree_extract(hammer_cursor_t cursor, int flags)
754 {
755         hammer_node_ondisk_t node;
756         hammer_btree_elm_t elm;
757         hammer_off_t data_off;
758         hammer_mount_t hmp;
759         int32_t data_len;
760         int error;
761
762         /*
763          * The case where the data reference resolves to the same buffer
764          * as the record reference must be handled.
765          */
766         node = cursor->node->ondisk;
767         elm = &node->elms[cursor->index];
768         cursor->data = NULL;
769         hmp = cursor->node->hmp;
770
771         /*
772          * There is nothing to extract for an internal element.
773          */
774         if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
775                 return(EINVAL);
776
777         /*
778          * Only record types have data.
779          */
780         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
781         cursor->leaf = &elm->leaf;
782
783         if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
784                 return(0);
785         if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
786                 return(0);
787         data_off = elm->leaf.data_offset;
788         data_len = elm->leaf.data_len;
789         if (data_off == 0)
790                 return(0);
791
792         /*
793          * Load the data
794          */
795         KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
796         cursor->data = hammer_bread_ext(hmp, data_off, data_len,
797                                         &error, &cursor->data_buffer);
798
799         /*
800          * Mark the data buffer as not being meta-data if it isn't
801          * meta-data (sometimes bulk data is accessed via a volume
802          * block device).
803          */
804         if (error == 0) {
805                 switch(elm->leaf.base.rec_type) {
806                 case HAMMER_RECTYPE_DATA:
807                 case HAMMER_RECTYPE_DB:
808                         if ((data_off & HAMMER_ZONE_LARGE_DATA) == 0)
809                                 break;
810                         if (hammer_double_buffer == 0 ||
811                             (cursor->flags & HAMMER_CURSOR_NOSWAPCACHE)) {
812                                 hammer_io_notmeta(cursor->data_buffer);
813                         }
814                         break;
815                 default:
816                         break;
817                 }
818         }
819
820         /*
821          * Deal with CRC errors on the extracted data.
822          */
823         if (error == 0 &&
824             hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0) {
825                 kprintf("CRC DATA @ %016llx/%d FAILED\n",
826                         (long long)elm->leaf.data_offset, elm->leaf.data_len);
827                 if (hammer_debug_critical)
828                         Debugger("CRC FAILED: DATA");
829                 if (cursor->trans->flags & HAMMER_TRANSF_CRCDOM)
830                         error = EDOM;   /* less critical (mirroring) */
831                 else
832                         error = EIO;    /* critical */
833         }
834         return(error);
835 }
836
837
838 /*
839  * Insert a leaf element into the B-Tree at the current cursor position.
840  * The cursor is positioned such that the element at and beyond the cursor
841  * are shifted to make room for the new record.
842  *
843  * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
844  * flag set and that call must return ENOENT before this function can be
845  * called.
846  *
847  * The caller may depend on the cursor's exclusive lock after return to
848  * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
849  *
850  * ENOSPC is returned if there is no room to insert a new record.
851  */
852 int
853 hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
854                     int *doprop)
855 {
856         hammer_node_ondisk_t node;
857         int i;
858         int error;
859
860         *doprop = 0;
861         if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
862                 return(error);
863         ++hammer_stats_btree_inserts;
864
865         /*
866          * Insert the element at the leaf node and update the count in the
867          * parent.  It is possible for parent to be NULL, indicating that
868          * the filesystem's ROOT B-Tree node is a leaf itself, which is
869          * possible.  The root inode can never be deleted so the leaf should
870          * never be empty.
871          *
872          * Remember that the right-hand boundary is not included in the
873          * count.
874          */
875         hammer_modify_node_all(cursor->trans, cursor->node);
876         node = cursor->node->ondisk;
877         i = cursor->index;
878         KKASSERT(elm->base.btype != 0);
879         KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
880         KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
881         if (i != node->count) {
882                 bcopy(&node->elms[i], &node->elms[i+1],
883                       (node->count - i) * sizeof(*elm));
884         }
885         node->elms[i].leaf = *elm;
886         ++node->count;
887         hammer_cursor_inserted_element(cursor->node, i);
888
889         /*
890          * Update the leaf node's aggregate mirror_tid for mirroring
891          * support.
892          */
893         if (node->mirror_tid < elm->base.delete_tid) {
894                 node->mirror_tid = elm->base.delete_tid;
895                 *doprop = 1;
896         }
897         if (node->mirror_tid < elm->base.create_tid) {
898                 node->mirror_tid = elm->base.create_tid;
899                 *doprop = 1;
900         }
901         hammer_modify_node_done(cursor->node);
902
903         /*
904          * Debugging sanity checks.
905          */
906         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
907         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
908         if (i) {
909                 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
910         }
911         if (i != node->count - 1)
912                 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
913
914         return(0);
915 }
916
917 /*
918  * Delete a record from the B-Tree at the current cursor position.
919  * The cursor is positioned such that the current element is the one
920  * to be deleted.
921  *
922  * On return the cursor will be positioned after the deleted element and
923  * MAY point to an internal node.  It will be suitable for the continuation
924  * of an iteration but not for an insertion or deletion.
925  *
926  * Deletions will attempt to partially rebalance the B-Tree in an upward
927  * direction, but will terminate rather then deadlock.  Empty internal nodes
928  * are never allowed by a deletion which deadlocks may end up giving us an
929  * empty leaf.  The pruner will clean up and rebalance the tree.
930  *
931  * This function can return EDEADLK, requiring the caller to retry the
932  * operation after clearing the deadlock.
933  */
934 int
935 hammer_btree_delete(hammer_cursor_t cursor)
936 {
937         hammer_node_ondisk_t ondisk;
938         hammer_node_t node;
939         hammer_node_t parent;
940         int error;
941         int i;
942
943         KKASSERT (cursor->trans->sync_lock_refs > 0);
944         if ((error = hammer_cursor_upgrade(cursor)) != 0)
945                 return(error);
946         ++hammer_stats_btree_deletes;
947
948         /*
949          * Delete the element from the leaf node. 
950          *
951          * Remember that leaf nodes do not have boundaries.
952          */
953         node = cursor->node;
954         ondisk = node->ondisk;
955         i = cursor->index;
956
957         KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
958         KKASSERT(i >= 0 && i < ondisk->count);
959         hammer_modify_node_all(cursor->trans, node);
960         if (i + 1 != ondisk->count) {
961                 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
962                       (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
963         }
964         --ondisk->count;
965         hammer_modify_node_done(node);
966         hammer_cursor_deleted_element(node, i);
967
968         /*
969          * Validate local parent
970          */
971         if (ondisk->parent) {
972                 parent = cursor->parent;
973
974                 KKASSERT(parent != NULL);
975                 KKASSERT(parent->node_offset == ondisk->parent);
976         }
977
978         /*
979          * If the leaf becomes empty it must be detached from the parent,
980          * potentially recursing through to the filesystem root.
981          *
982          * This may reposition the cursor at one of the parent's of the
983          * current node.
984          *
985          * Ignore deadlock errors, that simply means that btree_remove
986          * was unable to recurse and had to leave us with an empty leaf. 
987          */
988         KKASSERT(cursor->index <= ondisk->count);
989         if (ondisk->count == 0) {
990                 error = btree_remove(cursor);
991                 if (error == EDEADLK)
992                         error = 0;
993         } else {
994                 error = 0;
995         }
996         KKASSERT(cursor->parent == NULL ||
997                  cursor->parent_index < cursor->parent->ondisk->count);
998         return(error);
999 }
1000
1001 /*
1002  * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
1003  *
1004  * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
1005  *
1006  * The search can begin ANYWHERE in the B-Tree.  As a first step the search
1007  * iterates up the tree as necessary to properly position itself prior to
1008  * actually doing the sarch.
1009  * 
1010  * INSERTIONS: The search will split full nodes and leaves on its way down
1011  * and guarentee that the leaf it ends up on is not full.  If we run out
1012  * of space the search continues to the leaf (to position the cursor for
1013  * the spike), but ENOSPC is returned.
1014  *
1015  * The search is only guarenteed to end up on a leaf if an error code of 0
1016  * is returned, or if inserting and an error code of ENOENT is returned.
1017  * Otherwise it can stop at an internal node.  On success a search returns
1018  * a leaf node.
1019  *
1020  * COMPLEXITY WARNING!  This is the core B-Tree search code for the entire
1021  * filesystem, and it is not simple code.  Please note the following facts:
1022  *
1023  * - Internal node recursions have a boundary on the left AND right.  The
1024  *   right boundary is non-inclusive.  The create_tid is a generic part
1025  *   of the key for internal nodes.
1026  *
1027  * - Leaf nodes contain terminal elements only now.
1028  *
1029  * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
1030  *   historical search.  ASOF and INSERT are mutually exclusive.  When
1031  *   doing an as-of lookup btree_search() checks for a right-edge boundary
1032  *   case.  If while recursing down the left-edge differs from the key
1033  *   by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
1034  *   with cursor->create_check.  This is used by btree_lookup() to iterate.
1035  *   The iteration backwards because as-of searches can wind up going
1036  *   down the wrong branch of the B-Tree.
1037  */
1038 static 
1039 int
1040 btree_search(hammer_cursor_t cursor, int flags)
1041 {
1042         hammer_node_ondisk_t node;
1043         hammer_btree_elm_t elm;
1044         int error;
1045         int enospc = 0;
1046         int i;
1047         int r;
1048         int s;
1049
1050         flags |= cursor->flags;
1051         ++hammer_stats_btree_searches;
1052
1053         if (hammer_debug_btree) {
1054                 kprintf("SEARCH   %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
1055                         (long long)cursor->node->node_offset,
1056                         cursor->index,
1057                         (long long)cursor->key_beg.obj_id,
1058                         cursor->key_beg.rec_type,
1059                         (long long)cursor->key_beg.key,
1060                         (long long)cursor->key_beg.create_tid,
1061                         cursor->key_beg.localization, 
1062                         curthread
1063                 );
1064                 if (cursor->parent)
1065                     kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
1066                         (long long)cursor->parent->node_offset,
1067                         cursor->parent_index,
1068                         (long long)cursor->left_bound->obj_id,
1069                         (long long)cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
1070                         (long long)cursor->right_bound->obj_id,
1071                         (long long)cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
1072                         cursor->left_bound,
1073                         &cursor->parent->ondisk->elms[cursor->parent_index],
1074                         cursor->right_bound,
1075                         &cursor->parent->ondisk->elms[cursor->parent_index+1]
1076                     );
1077         }
1078
1079         /*
1080          * Move our cursor up the tree until we find a node whos range covers
1081          * the key we are trying to locate.
1082          *
1083          * The left bound is inclusive, the right bound is non-inclusive.
1084          * It is ok to cursor up too far.
1085          */
1086         for (;;) {
1087                 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
1088                 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
1089                 if (r >= 0 && s < 0)
1090                         break;
1091                 KKASSERT(cursor->parent);
1092                 ++hammer_stats_btree_iterations;
1093                 error = hammer_cursor_up(cursor);
1094                 if (error)
1095                         goto done;
1096         }
1097
1098         /*
1099          * The delete-checks below are based on node, not parent.  Set the
1100          * initial delete-check based on the parent.
1101          */
1102         if (r == 1) {
1103                 KKASSERT(cursor->left_bound->create_tid != 1);
1104                 cursor->create_check = cursor->left_bound->create_tid - 1;
1105                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1106         }
1107
1108         /*
1109          * We better have ended up with a node somewhere.
1110          */
1111         KKASSERT(cursor->node != NULL);
1112
1113         /*
1114          * If we are inserting we can't start at a full node if the parent
1115          * is also full (because there is no way to split the node),
1116          * continue running up the tree until the requirement is satisfied
1117          * or we hit the root of the filesystem.
1118          *
1119          * (If inserting we aren't doing an as-of search so we don't have
1120          *  to worry about create_check).
1121          */
1122         while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1123                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1124                         if (btree_node_is_full(cursor->node->ondisk) == 0)
1125                                 break;
1126                 } else {
1127                         if (btree_node_is_full(cursor->node->ondisk) ==0)
1128                                 break;
1129                 }
1130                 if (cursor->node->ondisk->parent == 0 ||
1131                     cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
1132                         break;
1133                 }
1134                 ++hammer_stats_btree_iterations;
1135                 error = hammer_cursor_up(cursor);
1136                 /* node may have become stale */
1137                 if (error)
1138                         goto done;
1139         }
1140
1141         /*
1142          * Push down through internal nodes to locate the requested key.
1143          */
1144         node = cursor->node->ondisk;
1145         while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
1146                 /*
1147                  * Scan the node to find the subtree index to push down into.
1148                  * We go one-past, then back-up.
1149                  *
1150                  * We must proactively remove deleted elements which may
1151                  * have been left over from a deadlocked btree_remove().
1152                  *
1153                  * The left and right boundaries are included in the loop
1154                  * in order to detect edge cases.
1155                  *
1156                  * If the separator only differs by create_tid (r == 1)
1157                  * and we are doing an as-of search, we may end up going
1158                  * down a branch to the left of the one containing the
1159                  * desired key.  This requires numerous special cases.
1160                  */
1161                 ++hammer_stats_btree_iterations;
1162                 if (hammer_debug_btree) {
1163                         kprintf("SEARCH-I %016llx count=%d\n",
1164                                 (long long)cursor->node->node_offset,
1165                                 node->count);
1166                 }
1167
1168                 /*
1169                  * Try to shortcut the search before dropping into the
1170                  * linear loop.  Locate the first node where r <= 1.
1171                  */
1172                 i = hammer_btree_search_node(&cursor->key_beg, node);
1173                 while (i <= node->count) {
1174                         ++hammer_stats_btree_elements;
1175                         elm = &node->elms[i];
1176                         r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
1177                         if (hammer_debug_btree > 2) {
1178                                 kprintf(" IELM %p %d r=%d\n",
1179                                         &node->elms[i], i, r);
1180                         }
1181                         if (r < 0)
1182                                 break;
1183                         if (r == 1) {
1184                                 KKASSERT(elm->base.create_tid != 1);
1185                                 cursor->create_check = elm->base.create_tid - 1;
1186                                 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
1187                         }
1188                         ++i;
1189                 }
1190                 if (hammer_debug_btree) {
1191                         kprintf("SEARCH-I preI=%d/%d r=%d\n",
1192                                 i, node->count, r);
1193                 }
1194
1195                 /*
1196                  * These cases occur when the parent's idea of the boundary
1197                  * is wider then the child's idea of the boundary, and
1198                  * require special handling.  If not inserting we can
1199                  * terminate the search early for these cases but the
1200                  * child's boundaries cannot be unconditionally modified.
1201                  */
1202                 if (i == 0) {
1203                         /*
1204                          * If i == 0 the search terminated to the LEFT of the
1205                          * left_boundary but to the RIGHT of the parent's left
1206                          * boundary.
1207                          */
1208                         u_int8_t save;
1209
1210                         elm = &node->elms[0];
1211
1212                         /*
1213                          * If we aren't inserting we can stop here.
1214                          */
1215                         if ((flags & (HAMMER_CURSOR_INSERT |
1216                                       HAMMER_CURSOR_PRUNING)) == 0) {
1217                                 cursor->index = 0;
1218                                 return(ENOENT);
1219                         }
1220
1221                         /*
1222                          * Correct a left-hand boundary mismatch.
1223                          *
1224                          * We can only do this if we can upgrade the lock,
1225                          * and synchronized as a background cursor (i.e.
1226                          * inserting or pruning).
1227                          *
1228                          * WARNING: We can only do this if inserting, i.e.
1229                          * we are running on the backend.
1230                          */
1231                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1232                                 return(error);
1233                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1234                         hammer_modify_node_field(cursor->trans, cursor->node,
1235                                                  elms[0]);
1236                         save = node->elms[0].base.btype;
1237                         node->elms[0].base = *cursor->left_bound;
1238                         node->elms[0].base.btype = save;
1239                         hammer_modify_node_done(cursor->node);
1240                 } else if (i == node->count + 1) {
1241                         /*
1242                          * If i == node->count + 1 the search terminated to
1243                          * the RIGHT of the right boundary but to the LEFT
1244                          * of the parent's right boundary.  If we aren't
1245                          * inserting we can stop here.
1246                          *
1247                          * Note that the last element in this case is
1248                          * elms[i-2] prior to adjustments to 'i'.
1249                          */
1250                         --i;
1251                         if ((flags & (HAMMER_CURSOR_INSERT |
1252                                       HAMMER_CURSOR_PRUNING)) == 0) {
1253                                 cursor->index = i;
1254                                 return (ENOENT);
1255                         }
1256
1257                         /*
1258                          * Correct a right-hand boundary mismatch.
1259                          * (actual push-down record is i-2 prior to
1260                          * adjustments to i).
1261                          *
1262                          * We can only do this if we can upgrade the lock,
1263                          * and synchronized as a background cursor (i.e.
1264                          * inserting or pruning).
1265                          *
1266                          * WARNING: We can only do this if inserting, i.e.
1267                          * we are running on the backend.
1268                          */
1269                         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1270                                 return(error);
1271                         elm = &node->elms[i];
1272                         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1273                         hammer_modify_node(cursor->trans, cursor->node,
1274                                            &elm->base, sizeof(elm->base));
1275                         elm->base = *cursor->right_bound;
1276                         hammer_modify_node_done(cursor->node);
1277                         --i;
1278                 } else {
1279                         /*
1280                          * The push-down index is now i - 1.  If we had
1281                          * terminated on the right boundary this will point
1282                          * us at the last element.
1283                          */
1284                         --i;
1285                 }
1286                 cursor->index = i;
1287                 elm = &node->elms[i];
1288
1289                 if (hammer_debug_btree) {
1290                         kprintf("RESULT-I %016llx[%d] %016llx %02x "
1291                                 "key=%016llx cre=%016llx lo=%02x\n",
1292                                 (long long)cursor->node->node_offset,
1293                                 i,
1294                                 (long long)elm->internal.base.obj_id,
1295                                 elm->internal.base.rec_type,
1296                                 (long long)elm->internal.base.key,
1297                                 (long long)elm->internal.base.create_tid,
1298                                 elm->internal.base.localization
1299                         );
1300                 }
1301
1302                 /*
1303                  * We better have a valid subtree offset.
1304                  */
1305                 KKASSERT(elm->internal.subtree_offset != 0);
1306
1307                 /*
1308                  * Handle insertion and deletion requirements.
1309                  *
1310                  * If inserting split full nodes.  The split code will
1311                  * adjust cursor->node and cursor->index if the current
1312                  * index winds up in the new node.
1313                  *
1314                  * If inserting and a left or right edge case was detected,
1315                  * we cannot correct the left or right boundary and must
1316                  * prepend and append an empty leaf node in order to make
1317                  * the boundary correction.
1318                  *
1319                  * If we run out of space we set enospc and continue on
1320                  * to a leaf to provide the spike code with a good point
1321                  * of entry.
1322                  */
1323                 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
1324                         if (btree_node_is_full(node)) {
1325                                 error = btree_split_internal(cursor);
1326                                 if (error) {
1327                                         if (error != ENOSPC)
1328                                                 goto done;
1329                                         enospc = 1;
1330                                 }
1331                                 /*
1332                                  * reload stale pointers
1333                                  */
1334                                 i = cursor->index;
1335                                 node = cursor->node->ondisk;
1336                         }
1337                 }
1338
1339                 /*
1340                  * Push down (push into new node, existing node becomes
1341                  * the parent) and continue the search.
1342                  */
1343                 error = hammer_cursor_down(cursor);
1344                 /* node may have become stale */
1345                 if (error)
1346                         goto done;
1347                 node = cursor->node->ondisk;
1348         }
1349
1350         /*
1351          * We are at a leaf, do a linear search of the key array.
1352          *
1353          * On success the index is set to the matching element and 0
1354          * is returned.
1355          *
1356          * On failure the index is set to the insertion point and ENOENT
1357          * is returned.
1358          *
1359          * Boundaries are not stored in leaf nodes, so the index can wind
1360          * up to the left of element 0 (index == 0) or past the end of
1361          * the array (index == node->count).  It is also possible that the
1362          * leaf might be empty.
1363          */
1364         ++hammer_stats_btree_iterations;
1365         KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
1366         KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
1367         if (hammer_debug_btree) {
1368                 kprintf("SEARCH-L %016llx count=%d\n",
1369                         (long long)cursor->node->node_offset,
1370                         node->count);
1371         }
1372
1373         /*
1374          * Try to shortcut the search before dropping into the
1375          * linear loop.  Locate the first node where r <= 1.
1376          */
1377         i = hammer_btree_search_node(&cursor->key_beg, node);
1378         while (i < node->count) {
1379                 ++hammer_stats_btree_elements;
1380                 elm = &node->elms[i];
1381
1382                 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
1383
1384                 if (hammer_debug_btree > 1)
1385                         kprintf("  ELM %p %d r=%d\n", &node->elms[i], i, r);
1386
1387                 /*
1388                  * We are at a record element.  Stop if we've flipped past
1389                  * key_beg, not counting the create_tid test.  Allow the
1390                  * r == 1 case (key_beg > element but differs only by its
1391                  * create_tid) to fall through to the AS-OF check.
1392                  */
1393                 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1394
1395                 if (r < 0)
1396                         goto failed;
1397                 if (r > 1) {
1398                         ++i;
1399                         continue;
1400                 }
1401
1402                 /*
1403                  * Check our as-of timestamp against the element.
1404                  */
1405                 if (flags & HAMMER_CURSOR_ASOF) {
1406                         if (hammer_btree_chkts(cursor->asof,
1407                                                &node->elms[i].base) != 0) {
1408                                 ++i;
1409                                 continue;
1410                         }
1411                         /* success */
1412                 } else {
1413                         if (r > 0) {    /* can only be +1 */
1414                                 ++i;
1415                                 continue;
1416                         }
1417                         /* success */
1418                 }
1419                 cursor->index = i;
1420                 error = 0;
1421                 if (hammer_debug_btree) {
1422                         kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1423                                 (long long)cursor->node->node_offset, i);
1424                 }
1425                 goto done;
1426         }
1427
1428         /*
1429          * The search of the leaf node failed.  i is the insertion point.
1430          */
1431 failed:
1432         if (hammer_debug_btree) {
1433                 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1434                         (long long)cursor->node->node_offset, i);
1435         }
1436
1437         /*
1438          * No exact match was found, i is now at the insertion point.
1439          *
1440          * If inserting split a full leaf before returning.  This
1441          * may have the side effect of adjusting cursor->node and
1442          * cursor->index.
1443          */
1444         cursor->index = i;
1445         if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
1446              btree_node_is_full(node)) {
1447                 error = btree_split_leaf(cursor);
1448                 if (error) {
1449                         if (error != ENOSPC)
1450                                 goto done;
1451                         enospc = 1;
1452                 }
1453                 /*
1454                  * reload stale pointers
1455                  */
1456                 /* NOT USED
1457                 i = cursor->index;
1458                 node = &cursor->node->internal;
1459                 */
1460         }
1461
1462         /*
1463          * We reached a leaf but did not find the key we were looking for.
1464          * If this is an insert we will be properly positioned for an insert
1465          * (ENOENT) or spike (ENOSPC) operation.
1466          */
1467         error = enospc ? ENOSPC : ENOENT;
1468 done:
1469         return(error);
1470 }
1471
1472 /*
1473  * Heuristical search for the first element whos comparison is <= 1.  May
1474  * return an index whos compare result is > 1 but may only return an index
1475  * whos compare result is <= 1 if it is the first element with that result.
1476  */
1477 int
1478 hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1479 {
1480         int b;
1481         int s;
1482         int i;
1483         int r;
1484
1485         /*
1486          * Don't bother if the node does not have very many elements
1487          */
1488         b = 0;
1489         s = node->count;
1490         while (s - b > 4) {
1491                 i = b + (s - b) / 2;
1492                 ++hammer_stats_btree_elements;
1493                 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1494                 if (r <= 1) {
1495                         s = i;
1496                 } else {
1497                         b = i;
1498                 }
1499         }
1500         return(b);
1501 }
1502
1503
1504 /************************************************************************
1505  *                         SPLITTING AND MERGING                        *
1506  ************************************************************************
1507  *
1508  * These routines do all the dirty work required to split and merge nodes.
1509  */
1510
1511 /*
1512  * Split an internal node into two nodes and move the separator at the split
1513  * point to the parent.
1514  *
1515  * (cursor->node, cursor->index) indicates the element the caller intends
1516  * to push into.  We will adjust node and index if that element winds
1517  * up in the split node.
1518  *
1519  * If we are at the root of the filesystem a new root must be created with
1520  * two elements, one pointing to the original root and one pointing to the
1521  * newly allocated split node.
1522  */
1523 static
1524 int
1525 btree_split_internal(hammer_cursor_t cursor)
1526 {
1527         hammer_node_ondisk_t ondisk;
1528         hammer_node_t node;
1529         hammer_node_t parent;
1530         hammer_node_t new_node;
1531         hammer_btree_elm_t elm;
1532         hammer_btree_elm_t parent_elm;
1533         struct hammer_node_lock lockroot;
1534         hammer_mount_t hmp = cursor->trans->hmp;
1535         hammer_off_t hint;
1536         int parent_index;
1537         int made_root;
1538         int split;
1539         int error;
1540         int i;
1541         const int esize = sizeof(*elm);
1542
1543         hammer_node_lock_init(&lockroot, cursor->node);
1544         error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
1545         if (error)
1546                 goto done;
1547         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1548                 goto done;
1549         ++hammer_stats_btree_splits;
1550
1551         /* 
1552          * Calculate the split point.  If the insertion point is at the
1553          * end of the leaf we adjust the split point significantly to the
1554          * right to try to optimize node fill and flag it.  If we hit
1555          * that same leaf again our heuristic failed and we don't try
1556          * to optimize node fill (it could lead to a degenerate case).
1557          */
1558         node = cursor->node;
1559         ondisk = node->ondisk;
1560         KKASSERT(ondisk->count > 4);
1561         if (cursor->index == ondisk->count &&
1562             (node->flags & HAMMER_NODE_NONLINEAR) == 0) {
1563                 split = (ondisk->count + 1) * 3 / 4;
1564                 node->flags |= HAMMER_NODE_NONLINEAR;
1565         } else {
1566                 /*
1567                  * We are splitting but elms[split] will be promoted to
1568                  * the parent, leaving the right hand node with one less
1569                  * element.  If the insertion point will be on the
1570                  * left-hand side adjust the split point to give the
1571                  * right hand side one additional node.
1572                  */
1573                 split = (ondisk->count + 1) / 2;
1574                 if (cursor->index <= split)
1575                         --split;
1576         }
1577
1578         /*
1579          * If we are at the root of the filesystem, create a new root node
1580          * with 1 element and split normally.  Avoid making major
1581          * modifications until we know the whole operation will work.
1582          */
1583         if (ondisk->parent == 0) {
1584                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1585                 if (parent == NULL)
1586                         goto done;
1587                 hammer_lock_ex(&parent->lock);
1588                 hammer_modify_node_noundo(cursor->trans, parent);
1589                 ondisk = parent->ondisk;
1590                 ondisk->count = 1;
1591                 ondisk->parent = 0;
1592                 ondisk->mirror_tid = node->ondisk->mirror_tid;
1593                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1594                 ondisk->elms[0].base = hmp->root_btree_beg;
1595                 ondisk->elms[0].base.btype = node->ondisk->type;
1596                 ondisk->elms[0].internal.subtree_offset = node->node_offset;
1597                 ondisk->elms[1].base = hmp->root_btree_end;
1598                 hammer_modify_node_done(parent);
1599                 /* ondisk->elms[1].base.btype - not used */
1600                 made_root = 1;
1601                 parent_index = 0;       /* index of current node in parent */
1602         } else {
1603                 made_root = 0;
1604                 parent = cursor->parent;
1605                 parent_index = cursor->parent_index;
1606         }
1607
1608         /*
1609          * Calculate a hint for the allocation of the new B-Tree node.
1610          * The most likely expansion is coming from the insertion point
1611          * at cursor->index, so try to localize the allocation of our
1612          * new node to accomodate that sub-tree.
1613          *
1614          * Use the right-most sub-tree when expandinging on the right edge.
1615          * This is a very common case when copying a directory tree.
1616          */
1617         if (cursor->index == ondisk->count)
1618                 hint = ondisk->elms[cursor->index - 1].internal.subtree_offset;
1619         else
1620                 hint = ondisk->elms[cursor->index].internal.subtree_offset;
1621
1622         /*
1623          * Split node into new_node at the split point.
1624          *
1625          *  B O O O P N N B     <-- P = node->elms[split] (index 4)
1626          *   0 1 2 3 4 5 6      <-- subtree indices
1627          *
1628          *       x x P x x
1629          *        s S S s  
1630          *         /   \
1631          *  B O O O B    B N N B        <--- inner boundary points are 'P'
1632          *   0 1 2 3      4 5 6  
1633          */
1634         new_node = hammer_alloc_btree(cursor->trans, 0, &error);
1635         if (new_node == NULL) {
1636                 if (made_root) {
1637                         hammer_unlock(&parent->lock);
1638                         hammer_delete_node(cursor->trans, parent);
1639                         hammer_rel_node(parent);
1640                 }
1641                 goto done;
1642         }
1643         hammer_lock_ex(&new_node->lock);
1644
1645         /*
1646          * Create the new node.  P becomes the left-hand boundary in the
1647          * new node.  Copy the right-hand boundary as well.
1648          *
1649          * elm is the new separator.
1650          */
1651         hammer_modify_node_noundo(cursor->trans, new_node);
1652         hammer_modify_node_all(cursor->trans, node);
1653         ondisk = node->ondisk;
1654         elm = &ondisk->elms[split];
1655         bcopy(elm, &new_node->ondisk->elms[0],
1656               (ondisk->count - split + 1) * esize);
1657         new_node->ondisk->count = ondisk->count - split;
1658         new_node->ondisk->parent = parent->node_offset;
1659         new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1660         new_node->ondisk->mirror_tid = ondisk->mirror_tid;
1661         KKASSERT(ondisk->type == new_node->ondisk->type);
1662         hammer_cursor_split_node(node, new_node, split);
1663
1664         /*
1665          * Cleanup the original node.  Elm (P) becomes the new boundary,
1666          * its subtree_offset was moved to the new node.  If we had created
1667          * a new root its parent pointer may have changed.
1668          */
1669         elm->internal.subtree_offset = 0;
1670         ondisk->count = split;
1671
1672         /*
1673          * Insert the separator into the parent, fixup the parent's
1674          * reference to the original node, and reference the new node.
1675          * The separator is P.
1676          *
1677          * Remember that base.count does not include the right-hand boundary.
1678          */
1679         hammer_modify_node_all(cursor->trans, parent);
1680         ondisk = parent->ondisk;
1681         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1682         parent_elm = &ondisk->elms[parent_index+1];
1683         bcopy(parent_elm, parent_elm + 1,
1684               (ondisk->count - parent_index) * esize);
1685         parent_elm->internal.base = elm->base;  /* separator P */
1686         parent_elm->internal.base.btype = new_node->ondisk->type;
1687         parent_elm->internal.subtree_offset = new_node->node_offset;
1688         parent_elm->internal.mirror_tid = new_node->ondisk->mirror_tid;
1689         ++ondisk->count;
1690         hammer_modify_node_done(parent);
1691         hammer_cursor_inserted_element(parent, parent_index + 1);
1692
1693         /*
1694          * The children of new_node need their parent pointer set to new_node.
1695          * The children have already been locked by
1696          * hammer_btree_lock_children().
1697          */
1698         for (i = 0; i < new_node->ondisk->count; ++i) {
1699                 elm = &new_node->ondisk->elms[i];
1700                 error = btree_set_parent(cursor->trans, new_node, elm);
1701                 if (error) {
1702                         panic("btree_split_internal: btree-fixup problem");
1703                 }
1704         }
1705         hammer_modify_node_done(new_node);
1706
1707         /*
1708          * The filesystem's root B-Tree pointer may have to be updated.
1709          */
1710         if (made_root) {
1711                 hammer_volume_t volume;
1712
1713                 volume = hammer_get_root_volume(hmp, &error);
1714                 KKASSERT(error == 0);
1715
1716                 hammer_modify_volume_field(cursor->trans, volume,
1717                                            vol0_btree_root);
1718                 volume->ondisk->vol0_btree_root = parent->node_offset;
1719                 hammer_modify_volume_done(volume);
1720                 node->ondisk->parent = parent->node_offset;
1721                 if (cursor->parent) {
1722                         hammer_unlock(&cursor->parent->lock);
1723                         hammer_rel_node(cursor->parent);
1724                 }
1725                 cursor->parent = parent;        /* lock'd and ref'd */
1726                 hammer_rel_volume(volume, 0);
1727         }
1728         hammer_modify_node_done(node);
1729
1730         /*
1731          * Ok, now adjust the cursor depending on which element the original
1732          * index was pointing at.  If we are >= the split point the push node
1733          * is now in the new node.
1734          *
1735          * NOTE: If we are at the split point itself we cannot stay with the
1736          * original node because the push index will point at the right-hand
1737          * boundary, which is illegal.
1738          *
1739          * NOTE: The cursor's parent or parent_index must be adjusted for
1740          * the case where a new parent (new root) was created, and the case
1741          * where the cursor is now pointing at the split node.
1742          */
1743         if (cursor->index >= split) {
1744                 cursor->parent_index = parent_index + 1;
1745                 cursor->index -= split;
1746                 hammer_unlock(&cursor->node->lock);
1747                 hammer_rel_node(cursor->node);
1748                 cursor->node = new_node;        /* locked and ref'd */
1749         } else {
1750                 cursor->parent_index = parent_index;
1751                 hammer_unlock(&new_node->lock);
1752                 hammer_rel_node(new_node);
1753         }
1754
1755         /*
1756          * Fixup left and right bounds
1757          */
1758         parent_elm = &parent->ondisk->elms[cursor->parent_index];
1759         cursor->left_bound = &parent_elm[0].internal.base;
1760         cursor->right_bound = &parent_elm[1].internal.base;
1761         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1762                  &cursor->node->ondisk->elms[0].internal.base) <= 0);
1763         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1764                  &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
1765
1766 done:
1767         hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
1768         hammer_cursor_downgrade(cursor);
1769         return (error);
1770 }
1771
1772 /*
1773  * Same as the above, but splits a full leaf node.
1774  *
1775  * This function
1776  */
1777 static
1778 int
1779 btree_split_leaf(hammer_cursor_t cursor)
1780 {
1781         hammer_node_ondisk_t ondisk;
1782         hammer_node_t parent;
1783         hammer_node_t leaf;
1784         hammer_mount_t hmp;
1785         hammer_node_t new_leaf;
1786         hammer_btree_elm_t elm;
1787         hammer_btree_elm_t parent_elm;
1788         hammer_base_elm_t mid_boundary;
1789         hammer_off_t hint;
1790         int parent_index;
1791         int made_root;
1792         int split;
1793         int error;
1794         const size_t esize = sizeof(*elm);
1795
1796         if ((error = hammer_cursor_upgrade(cursor)) != 0)
1797                 return(error);
1798         ++hammer_stats_btree_splits;
1799
1800         KKASSERT(hammer_btree_cmp(cursor->left_bound,
1801                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1802         KKASSERT(hammer_btree_cmp(cursor->right_bound,
1803                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1804
1805         /* 
1806          * Calculate the split point.  If the insertion point is at the
1807          * end of the leaf we adjust the split point significantly to the
1808          * right to try to optimize node fill and flag it.  If we hit
1809          * that same leaf again our heuristic failed and we don't try
1810          * to optimize node fill (it could lead to a degenerate case).
1811          *
1812          * Spikes are made up of two leaf elements which cannot be
1813          * safely split.
1814          */
1815         leaf = cursor->node;
1816         ondisk = leaf->ondisk;
1817         KKASSERT(ondisk->count > 4);
1818         if (cursor->index == ondisk->count &&
1819             (leaf->flags & HAMMER_NODE_NONLINEAR) == 0) {
1820                 split = (ondisk->count + 1) * 3 / 4;
1821                 leaf->flags |= HAMMER_NODE_NONLINEAR;
1822         } else {
1823                 split = (ondisk->count + 1) / 2;
1824         }
1825
1826 #if 0
1827         /*
1828          * If the insertion point is at the split point shift the
1829          * split point left so we don't have to worry about
1830          */
1831         if (cursor->index == split)
1832                 --split;
1833 #endif
1834         KKASSERT(split > 0 && split < ondisk->count);
1835
1836         error = 0;
1837         hmp = leaf->hmp;
1838
1839         elm = &ondisk->elms[split];
1840
1841         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1842         KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1843         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1844         KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1845
1846         /*
1847          * If we are at the root of the tree, create a new root node with
1848          * 1 element and split normally.  Avoid making major modifications
1849          * until we know the whole operation will work.
1850          */
1851         if (ondisk->parent == 0) {
1852                 parent = hammer_alloc_btree(cursor->trans, 0, &error);
1853                 if (parent == NULL)
1854                         goto done;
1855                 hammer_lock_ex(&parent->lock);
1856                 hammer_modify_node_noundo(cursor->trans, parent);
1857                 ondisk = parent->ondisk;
1858                 ondisk->count = 1;
1859                 ondisk->parent = 0;
1860                 ondisk->mirror_tid = leaf->ondisk->mirror_tid;
1861                 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1862                 ondisk->elms[0].base = hmp->root_btree_beg;
1863                 ondisk->elms[0].base.btype = leaf->ondisk->type;
1864                 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
1865                 ondisk->elms[1].base = hmp->root_btree_end;
1866                 /* ondisk->elms[1].base.btype = not used */
1867                 hammer_modify_node_done(parent);
1868                 made_root = 1;
1869                 parent_index = 0;       /* insertion point in parent */
1870         } else {
1871                 made_root = 0;
1872                 parent = cursor->parent;
1873                 parent_index = cursor->parent_index;
1874         }
1875
1876         /*
1877          * Calculate a hint for the allocation of the new B-Tree leaf node.
1878          * For now just try to localize it within the same bigblock as
1879          * the current leaf.
1880          *
1881          * If the insertion point is at the end of the leaf we recognize a
1882          * likely append sequence of some sort (data, meta-data, inodes,
1883          * whatever).  Set the hint to zero to allocate out of linear space
1884          * instead of trying to completely fill previously hinted space.
1885          *
1886          * This also sets the stage for recursive splits to localize using
1887          * the new space.
1888          */
1889         ondisk = leaf->ondisk;
1890         if (cursor->index == ondisk->count)
1891                 hint = 0;
1892         else
1893                 hint = leaf->node_offset;
1894
1895         /*
1896          * Split leaf into new_leaf at the split point.  Select a separator
1897          * value in-between the two leafs but with a bent towards the right
1898          * leaf since comparisons use an 'elm >= separator' inequality.
1899          *
1900          *  L L L L L L L L
1901          *
1902          *       x x P x x
1903          *        s S S s  
1904          *         /   \
1905          *  L L L L     L L L L
1906          */
1907         new_leaf = hammer_alloc_btree(cursor->trans, 0, &error);
1908         if (new_leaf == NULL) {
1909                 if (made_root) {
1910                         hammer_unlock(&parent->lock);
1911                         hammer_delete_node(cursor->trans, parent);
1912                         hammer_rel_node(parent);
1913                 }
1914                 goto done;
1915         }
1916         hammer_lock_ex(&new_leaf->lock);
1917
1918         /*
1919          * Create the new node and copy the leaf elements from the split 
1920          * point on to the new node.
1921          */
1922         hammer_modify_node_all(cursor->trans, leaf);
1923         hammer_modify_node_noundo(cursor->trans, new_leaf);
1924         ondisk = leaf->ondisk;
1925         elm = &ondisk->elms[split];
1926         bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1927         new_leaf->ondisk->count = ondisk->count - split;
1928         new_leaf->ondisk->parent = parent->node_offset;
1929         new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1930         new_leaf->ondisk->mirror_tid = ondisk->mirror_tid;
1931         KKASSERT(ondisk->type == new_leaf->ondisk->type);
1932         hammer_modify_node_done(new_leaf);
1933         hammer_cursor_split_node(leaf, new_leaf, split);
1934
1935         /*
1936          * Cleanup the original node.  Because this is a leaf node and
1937          * leaf nodes do not have a right-hand boundary, there
1938          * aren't any special edge cases to clean up.  We just fixup the
1939          * count.
1940          */
1941         ondisk->count = split;
1942
1943         /*
1944          * Insert the separator into the parent, fixup the parent's
1945          * reference to the original node, and reference the new node.
1946          * The separator is P.
1947          *
1948          * Remember that base.count does not include the right-hand boundary.
1949          * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1950          */
1951         hammer_modify_node_all(cursor->trans, parent);
1952         ondisk = parent->ondisk;
1953         KKASSERT(split != 0);
1954         KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
1955         parent_elm = &ondisk->elms[parent_index+1];
1956         bcopy(parent_elm, parent_elm + 1,
1957               (ondisk->count - parent_index) * esize);
1958
1959         hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
1960         parent_elm->internal.base.btype = new_leaf->ondisk->type;
1961         parent_elm->internal.subtree_offset = new_leaf->node_offset;
1962         parent_elm->internal.mirror_tid = new_leaf->ondisk->mirror_tid;
1963         mid_boundary = &parent_elm->base;
1964         ++ondisk->count;
1965         hammer_modify_node_done(parent);
1966         hammer_cursor_inserted_element(parent, parent_index + 1);
1967
1968         /*
1969          * The filesystem's root B-Tree pointer may have to be updated.
1970          */
1971         if (made_root) {
1972                 hammer_volume_t volume;
1973
1974                 volume = hammer_get_root_volume(hmp, &error);
1975                 KKASSERT(error == 0);
1976
1977                 hammer_modify_volume_field(cursor->trans, volume,
1978                                            vol0_btree_root);
1979                 volume->ondisk->vol0_btree_root = parent->node_offset;
1980                 hammer_modify_volume_done(volume);
1981                 leaf->ondisk->parent = parent->node_offset;
1982                 if (cursor->parent) {
1983                         hammer_unlock(&cursor->parent->lock);
1984                         hammer_rel_node(cursor->parent);
1985                 }
1986                 cursor->parent = parent;        /* lock'd and ref'd */
1987                 hammer_rel_volume(volume, 0);
1988         }
1989         hammer_modify_node_done(leaf);
1990
1991         /*
1992          * Ok, now adjust the cursor depending on which element the original
1993          * index was pointing at.  If we are >= the split point the push node
1994          * is now in the new node.
1995          *
1996          * NOTE: If we are at the split point itself we need to select the
1997          * old or new node based on where key_beg's insertion point will be.
1998          * If we pick the wrong side the inserted element will wind up in
1999          * the wrong leaf node and outside that node's bounds.
2000          */
2001         if (cursor->index > split ||
2002             (cursor->index == split &&
2003              hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
2004                 cursor->parent_index = parent_index + 1;
2005                 cursor->index -= split;
2006                 hammer_unlock(&cursor->node->lock);
2007                 hammer_rel_node(cursor->node);
2008                 cursor->node = new_leaf;
2009         } else {
2010                 cursor->parent_index = parent_index;
2011                 hammer_unlock(&new_leaf->lock);
2012                 hammer_rel_node(new_leaf);
2013         }
2014
2015         /*
2016          * Fixup left and right bounds
2017          */
2018         parent_elm = &parent->ondisk->elms[cursor->parent_index];
2019         cursor->left_bound = &parent_elm[0].internal.base;
2020         cursor->right_bound = &parent_elm[1].internal.base;
2021
2022         /*
2023          * Assert that the bounds are correct.
2024          */
2025         KKASSERT(hammer_btree_cmp(cursor->left_bound,
2026                  &cursor->node->ondisk->elms[0].leaf.base) <= 0);
2027         KKASSERT(hammer_btree_cmp(cursor->right_bound,
2028                  &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
2029         KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
2030         KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
2031
2032 done:
2033         hammer_cursor_downgrade(cursor);
2034         return (error);
2035 }
2036
2037 #if 0
2038
2039 /*
2040  * Recursively correct the right-hand boundary's create_tid to (tid) as
2041  * long as the rest of the key matches.  We have to recurse upward in
2042  * the tree as well as down the left side of each parent's right node.
2043  *
2044  * Return EDEADLK if we were only partially successful, forcing the caller
2045  * to try again.  The original cursor is not modified.  This routine can
2046  * also fail with EDEADLK if it is forced to throw away a portion of its
2047  * record history.
2048  *
2049  * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
2050  */
2051 struct hammer_rhb {
2052         TAILQ_ENTRY(hammer_rhb) entry;
2053         hammer_node_t   node;
2054         int             index;
2055 };
2056
2057 TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
2058
2059 int
2060 hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
2061 {
2062         struct hammer_mount *hmp;
2063         struct hammer_rhb_list rhb_list;
2064         hammer_base_elm_t elm;
2065         hammer_node_t orig_node;
2066         struct hammer_rhb *rhb;
2067         int orig_index;
2068         int error;
2069
2070         TAILQ_INIT(&rhb_list);
2071         hmp = cursor->trans->hmp;
2072
2073         /*
2074          * Save our position so we can restore it on return.  This also
2075          * gives us a stable 'elm'.
2076          */
2077         orig_node = cursor->node;
2078         hammer_ref_node(orig_node);
2079         hammer_lock_sh(&orig_node->lock);
2080         orig_index = cursor->index;
2081         elm = &orig_node->ondisk->elms[orig_index].base;
2082
2083         /*
2084          * Now build a list of parents going up, allocating a rhb
2085          * structure for each one.
2086          */
2087         while (cursor->parent) {
2088                 /*
2089                  * Stop if we no longer have any right-bounds to fix up
2090                  */
2091                 if (elm->obj_id != cursor->right_bound->obj_id ||
2092                     elm->rec_type != cursor->right_bound->rec_type ||
2093                     elm->key != cursor->right_bound->key) {
2094                         break;
2095                 }
2096
2097                 /*
2098                  * Stop if the right-hand bound's create_tid does not
2099                  * need to be corrected.
2100                  */
2101                 if (cursor->right_bound->create_tid >= tid)
2102                         break;
2103
2104                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2105                 rhb->node = cursor->parent;
2106                 rhb->index = cursor->parent_index;
2107                 hammer_ref_node(rhb->node);
2108                 hammer_lock_sh(&rhb->node->lock);
2109                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2110
2111                 hammer_cursor_up(cursor);
2112         }
2113
2114         /*
2115          * now safely adjust the right hand bound for each rhb.  This may
2116          * also require taking the right side of the tree and iterating down
2117          * ITS left side.
2118          */
2119         error = 0;
2120         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2121                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2122                 if (error)
2123                         break;
2124                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2125                 hammer_unlock(&rhb->node->lock);
2126                 hammer_rel_node(rhb->node);
2127                 kfree(rhb, hmp->m_misc);
2128
2129                 switch (cursor->node->ondisk->type) {
2130                 case HAMMER_BTREE_TYPE_INTERNAL:
2131                         /*
2132                          * Right-boundary for parent at internal node
2133                          * is one element to the right of the element whos
2134                          * right boundary needs adjusting.  We must then
2135                          * traverse down the left side correcting any left
2136                          * bounds (which may now be too far to the left).
2137                          */
2138                         ++cursor->index;
2139                         error = hammer_btree_correct_lhb(cursor, tid);
2140                         break;
2141                 default:
2142                         panic("hammer_btree_correct_rhb(): Bad node type");
2143                         error = EINVAL;
2144                         break;
2145                 }
2146         }
2147
2148         /*
2149          * Cleanup
2150          */
2151         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2152                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2153                 hammer_unlock(&rhb->node->lock);
2154                 hammer_rel_node(rhb->node);
2155                 kfree(rhb, hmp->m_misc);
2156         }
2157         error = hammer_cursor_seek(cursor, orig_node, orig_index);
2158         hammer_unlock(&orig_node->lock);
2159         hammer_rel_node(orig_node);
2160         return (error);
2161 }
2162
2163 /*
2164  * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
2165  * bound going downward starting at the current cursor position.
2166  *
2167  * This function does not restore the cursor after use.
2168  */
2169 int
2170 hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
2171 {
2172         struct hammer_rhb_list rhb_list;
2173         hammer_base_elm_t elm;
2174         hammer_base_elm_t cmp;
2175         struct hammer_rhb *rhb;
2176         struct hammer_mount *hmp;
2177         int error;
2178
2179         TAILQ_INIT(&rhb_list);
2180         hmp = cursor->trans->hmp;
2181
2182         cmp = &cursor->node->ondisk->elms[cursor->index].base;
2183
2184         /*
2185          * Record the node and traverse down the left-hand side for all
2186          * matching records needing a boundary correction.
2187          */
2188         error = 0;
2189         for (;;) {
2190                 rhb = kmalloc(sizeof(*rhb), hmp->m_misc, M_WAITOK|M_ZERO);
2191                 rhb->node = cursor->node;
2192                 rhb->index = cursor->index;
2193                 hammer_ref_node(rhb->node);
2194                 hammer_lock_sh(&rhb->node->lock);
2195                 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
2196
2197                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2198                         /*
2199                          * Nothing to traverse down if we are at the right
2200                          * boundary of an internal node.
2201                          */
2202                         if (cursor->index == cursor->node->ondisk->count)
2203                                 break;
2204                 } else {
2205                         elm = &cursor->node->ondisk->elms[cursor->index].base;
2206                         if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
2207                                 break;
2208                         panic("Illegal leaf record type %02x", elm->btype);
2209                 }
2210                 error = hammer_cursor_down(cursor);
2211                 if (error)
2212                         break;
2213
2214                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2215                 if (elm->obj_id != cmp->obj_id ||
2216                     elm->rec_type != cmp->rec_type ||
2217                     elm->key != cmp->key) {
2218                         break;
2219                 }
2220                 if (elm->create_tid >= tid)
2221                         break;
2222
2223         }
2224
2225         /*
2226          * Now we can safely adjust the left-hand boundary from the bottom-up.
2227          * The last element we remove from the list is the caller's right hand
2228          * boundary, which must also be adjusted.
2229          */
2230         while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2231                 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
2232                 if (error)
2233                         break;
2234                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2235                 hammer_unlock(&rhb->node->lock);
2236                 hammer_rel_node(rhb->node);
2237                 kfree(rhb, hmp->m_misc);
2238
2239                 elm = &cursor->node->ondisk->elms[cursor->index].base;
2240                 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2241                         hammer_modify_node(cursor->trans, cursor->node,
2242                                            &elm->create_tid,
2243                                            sizeof(elm->create_tid));
2244                         elm->create_tid = tid;
2245                         hammer_modify_node_done(cursor->node);
2246                 } else {
2247                         panic("hammer_btree_correct_lhb(): Bad element type");
2248                 }
2249         }
2250
2251         /*
2252          * Cleanup
2253          */
2254         while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
2255                 TAILQ_REMOVE(&rhb_list, rhb, entry);
2256                 hammer_unlock(&rhb->node->lock);
2257                 hammer_rel_node(rhb->node);
2258                 kfree(rhb, hmp->m_misc);
2259         }
2260         return (error);
2261 }
2262
2263 #endif
2264
2265 /*
2266  * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2267  * (cursor->node).  Returns 0 on success, EDEADLK if we could not complete
2268  * the operation due to a deadlock, or some other error.
2269  *
2270  * This routine is initially called with an empty leaf and may be
2271  * recursively called with single-element internal nodes.
2272  *
2273  * It should also be noted that when removing empty leaves we must be sure
2274  * to test and update mirror_tid because another thread may have deadlocked
2275  * against us (or someone) trying to propagate it up and cannot retry once
2276  * the node has been deleted.
2277  *
2278  * On return the cursor may end up pointing to an internal node, suitable
2279  * for further iteration but not for an immediate insertion or deletion.
2280  */
2281 static int
2282 btree_remove(hammer_cursor_t cursor)
2283 {
2284         hammer_node_ondisk_t ondisk;
2285         hammer_btree_elm_t elm;
2286         hammer_node_t node;
2287         hammer_node_t parent;
2288         const int esize = sizeof(*elm);
2289         int error;
2290
2291         node = cursor->node;
2292
2293         /*
2294          * When deleting the root of the filesystem convert it to
2295          * an empty leaf node.  Internal nodes cannot be empty.
2296          */
2297         ondisk = node->ondisk;
2298         if (ondisk->parent == 0) {
2299                 KKASSERT(cursor->parent == NULL);
2300                 hammer_modify_node_all(cursor->trans, node);
2301                 KKASSERT(ondisk == node->ondisk);
2302                 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2303                 ondisk->count = 0;
2304                 hammer_modify_node_done(node);
2305                 cursor->index = 0;
2306                 return(0);
2307         }
2308
2309         parent = cursor->parent;
2310
2311         /*
2312          * Attempt to remove the parent's reference to the child.  If the
2313          * parent would become empty we have to recurse.  If we fail we 
2314          * leave the parent pointing to an empty leaf node.
2315          *
2316          * We have to recurse successfully before we can delete the internal
2317          * node as it is illegal to have empty internal nodes.  Even though
2318          * the operation may be aborted we must still fixup any unlocked
2319          * cursors as if we had deleted the element prior to recursing
2320          * (by calling hammer_cursor_deleted_element()) so those cursors
2321          * are properly forced up the chain by the recursion.
2322          */
2323         if (parent->ondisk->count == 1) {
2324                 /*
2325                  * This special cursor_up_locked() call leaves the original
2326                  * node exclusively locked and referenced, leaves the
2327                  * original parent locked (as the new node), and locks the
2328                  * new parent.  It can return EDEADLK.
2329                  *
2330                  * We cannot call hammer_cursor_removed_node() until we are
2331                  * actually able to remove the node.  If we did then tracked
2332                  * cursors in the middle of iterations could be repointed
2333                  * to a parent node.  If this occurs they could end up
2334                  * scanning newly inserted records into the node (that could
2335                  * not be deleted) when they push down again.
2336                  *
2337                  * Due to the way the recursion works the final parent is left
2338                  * in cursor->parent after the recursion returns.  Each
2339                  * layer on the way back up is thus able to call
2340                  * hammer_cursor_removed_node() and 'jump' the node up to
2341                  * the (same) final parent.
2342                  *
2343                  * NOTE!  The local variable 'parent' is invalid after we
2344                  *        call hammer_cursor_up_locked().
2345                  */
2346                 error = hammer_cursor_up_locked(cursor);
2347                 parent = NULL;
2348
2349                 if (error == 0) {
2350                         hammer_cursor_deleted_element(cursor->node, 0);
2351                         error = btree_remove(cursor);
2352                         if (error == 0) {
2353                                 KKASSERT(node != cursor->node);
2354                                 hammer_cursor_removed_node(
2355                                         node, cursor->node,
2356                                         cursor->index);
2357                                 hammer_modify_node_all(cursor->trans, node);
2358                                 ondisk = node->ondisk;
2359                                 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2360                                 ondisk->count = 0;
2361                                 hammer_modify_node_done(node);
2362                                 hammer_flush_node(node, 0);
2363                                 hammer_delete_node(cursor->trans, node);
2364                         } else {
2365                                 /*
2366                                  * Defer parent removal because we could not
2367                                  * get the lock, just let the leaf remain
2368                                  * empty.
2369                                  */
2370                                 /**/
2371                         }
2372                         hammer_unlock(&node->lock);
2373                         hammer_rel_node(node);
2374                 } else {
2375                         /*
2376                          * Defer parent removal because we could not
2377                          * get the lock, just let the leaf remain
2378                          * empty.
2379                          */
2380                         /**/
2381                 }
2382         } else {
2383                 KKASSERT(parent->ondisk->count > 1);
2384
2385                 hammer_modify_node_all(cursor->trans, parent);
2386                 ondisk = parent->ondisk;
2387                 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2388
2389                 elm = &ondisk->elms[cursor->parent_index];
2390                 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2391                 KKASSERT(ondisk->count > 0);
2392
2393                 /*
2394                  * We must retain the highest mirror_tid.  The deleted
2395                  * range is now encompassed by the element to the left.
2396                  * If we are already at the left edge the new left edge
2397                  * inherits mirror_tid.
2398                  *
2399                  * Note that bounds of the parent to our parent may create
2400                  * a gap to the left of our left-most node or to the right
2401                  * of our right-most node.  The gap is silently included
2402                  * in the mirror_tid's area of effect from the point of view
2403                  * of the scan.
2404                  */
2405                 if (cursor->parent_index) {
2406                         if (elm[-1].internal.mirror_tid <
2407                             elm[0].internal.mirror_tid) {
2408                                 elm[-1].internal.mirror_tid =
2409                                     elm[0].internal.mirror_tid;
2410                         }
2411                 } else {
2412                         if (elm[1].internal.mirror_tid <
2413                             elm[0].internal.mirror_tid) {
2414                                 elm[1].internal.mirror_tid =
2415                                     elm[0].internal.mirror_tid;
2416                         }
2417                 }
2418
2419                 /*
2420                  * Delete the subtree reference in the parent.  Include
2421                  * boundary element at end.
2422                  */
2423                 bcopy(&elm[1], &elm[0],
2424                       (ondisk->count - cursor->parent_index) * esize);
2425                 --ondisk->count;
2426                 hammer_modify_node_done(parent);
2427                 hammer_cursor_removed_node(node, parent, cursor->parent_index);
2428                 hammer_cursor_deleted_element(parent, cursor->parent_index);
2429                 hammer_flush_node(node, 0);
2430                 hammer_delete_node(cursor->trans, node);
2431
2432                 /*
2433                  * cursor->node is invalid, cursor up to make the cursor
2434                  * valid again.  We have to flag the condition in case
2435                  * another thread wiggles an insertion in during an
2436                  * iteration.
2437                  */
2438                 cursor->flags |= HAMMER_CURSOR_ITERATE_CHECK;
2439                 error = hammer_cursor_up(cursor);
2440         }
2441         return (error);
2442 }
2443
2444 /*
2445  * Propagate cursor->trans->tid up the B-Tree starting at the current
2446  * cursor position using pseudofs info gleaned from the passed inode.
2447  *
2448  * The passed inode has no relationship to the cursor position other
2449  * then being in the same pseudofs as the insertion or deletion we
2450  * are propagating the mirror_tid for.
2451  *
2452  * WARNING!  Because we push and pop the passed cursor, it may be
2453  *           modified by other B-Tree operations while it is unlocked
2454  *           and things like the node & leaf pointers, and indexes might
2455  *           change.
2456  */
2457 void
2458 hammer_btree_do_propagation(hammer_cursor_t cursor,
2459                             hammer_pseudofs_inmem_t pfsm,
2460                             hammer_btree_leaf_elm_t leaf)
2461 {
2462         hammer_cursor_t ncursor;
2463         hammer_tid_t mirror_tid;
2464         int error;
2465
2466         /*
2467          * We do not propagate a mirror_tid if the filesystem was mounted
2468          * in no-mirror mode.
2469          */
2470         if (cursor->trans->hmp->master_id < 0)
2471                 return;
2472
2473         /*
2474          * This is a bit of a hack because we cannot deadlock or return
2475          * EDEADLK here.  The related operation has already completed and
2476          * we must propagate the mirror_tid now regardless.
2477          *
2478          * Generate a new cursor which inherits the original's locks and
2479          * unlock the original.  Use the new cursor to propagate the
2480          * mirror_tid.  Then clean up the new cursor and reacquire locks
2481          * on the original.
2482          *
2483          * hammer_dup_cursor() cannot dup locks.  The dup inherits the
2484          * original's locks and the original is tracked and must be
2485          * re-locked.
2486          */
2487         mirror_tid = cursor->node->ondisk->mirror_tid;
2488         KKASSERT(mirror_tid != 0);
2489         ncursor = hammer_push_cursor(cursor);
2490         error = hammer_btree_mirror_propagate(ncursor, mirror_tid);
2491         KKASSERT(error == 0);
2492         hammer_pop_cursor(cursor, ncursor);
2493         /* WARNING: cursor's leaf pointer may change after pop */
2494 }
2495
2496
2497 /*
2498  * Propagate a mirror TID update upwards through the B-Tree to the root.
2499  *
2500  * A locked internal node must be passed in.  The node will remain locked
2501  * on return.
2502  *
2503  * This function syncs mirror_tid at the specified internal node's element,
2504  * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2505  */
2506 static int
2507 hammer_btree_mirror_propagate(hammer_cursor_t cursor, hammer_tid_t mirror_tid)
2508 {
2509         hammer_btree_internal_elm_t elm;
2510         hammer_node_t node;
2511         int error;
2512
2513         for (;;) {
2514                 error = hammer_cursor_up(cursor);
2515                 if (error == 0)
2516                         error = hammer_cursor_upgrade(cursor);
2517
2518                 /*
2519                  * We can ignore HAMMER_CURSOR_ITERATE_CHECK, the
2520                  * cursor will still be properly positioned for
2521                  * mirror propagation, just not for iterations.
2522                  */
2523                 while (error == EDEADLK) {
2524                         hammer_recover_cursor(cursor);
2525                         error = hammer_cursor_upgrade(cursor);
2526                 }
2527                 if (error)
2528                         break;
2529
2530                 /*
2531                  * If the cursor deadlocked it could end up at a leaf
2532                  * after we lost the lock.
2533                  */
2534                 node = cursor->node;
2535                 if (node->ondisk->type != HAMMER_BTREE_TYPE_INTERNAL)
2536                         continue;
2537
2538                 /*
2539                  * Adjust the node's element
2540                  */
2541                 elm = &node->ondisk->elms[cursor->index].internal;
2542                 if (elm->mirror_tid >= mirror_tid)
2543                         break;
2544                 hammer_modify_node(cursor->trans, node, &elm->mirror_tid,
2545                                    sizeof(elm->mirror_tid));
2546                 elm->mirror_tid = mirror_tid;
2547                 hammer_modify_node_done(node);
2548                 if (hammer_debug_general & 0x0002) {
2549                         kprintf("mirror_propagate: propagate "
2550                                 "%016llx @%016llx:%d\n",
2551                                 (long long)mirror_tid,
2552                                 (long long)node->node_offset,
2553                                 cursor->index);
2554                 }
2555
2556
2557                 /*
2558                  * Adjust the node's mirror_tid aggregator
2559                  */
2560                 if (node->ondisk->mirror_tid >= mirror_tid)
2561                         return(0);
2562                 hammer_modify_node_field(cursor->trans, node, mirror_tid);
2563                 node->ondisk->mirror_tid = mirror_tid;
2564                 hammer_modify_node_done(node);
2565                 if (hammer_debug_general & 0x0002) {
2566                         kprintf("mirror_propagate: propagate "
2567                                 "%016llx @%016llx\n",
2568                                 (long long)mirror_tid,
2569                                 (long long)node->node_offset);
2570                 }
2571         }
2572         if (error == ENOENT)
2573                 error = 0;
2574         return(error);
2575 }
2576
2577 hammer_node_t
2578 hammer_btree_get_parent(hammer_transaction_t trans, hammer_node_t node,
2579                         int *parent_indexp, int *errorp, int try_exclusive)
2580 {
2581         hammer_node_t parent;
2582         hammer_btree_elm_t elm;
2583         int i;
2584
2585         /*
2586          * Get the node
2587          */
2588         parent = hammer_get_node(trans, node->ondisk->parent, 0, errorp);
2589         if (*errorp) {
2590                 KKASSERT(parent == NULL);
2591                 return(NULL);
2592         }
2593         KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2594
2595         /*
2596          * Lock the node
2597          */
2598         if (try_exclusive) {
2599                 if (hammer_lock_ex_try(&parent->lock)) {
2600                         hammer_rel_node(parent);
2601                         *errorp = EDEADLK;
2602                         return(NULL);
2603                 }
2604         } else {
2605                 hammer_lock_sh(&parent->lock);
2606         }
2607
2608         /*
2609          * Figure out which element in the parent is pointing to the
2610          * child.
2611          */
2612         if (node->ondisk->count) {
2613                 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2614                                              parent->ondisk);
2615         } else {
2616                 i = 0;
2617         }
2618         while (i < parent->ondisk->count) {
2619                 elm = &parent->ondisk->elms[i];
2620                 if (elm->internal.subtree_offset == node->node_offset)
2621                         break;
2622                 ++i;
2623         }
2624         if (i == parent->ondisk->count) {
2625                 hammer_unlock(&parent->lock);
2626                 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2627         }
2628         *parent_indexp = i;
2629         KKASSERT(*errorp == 0);
2630         return(parent);
2631 }
2632
2633 /*
2634  * The element (elm) has been moved to a new internal node (node).
2635  *
2636  * If the element represents a pointer to an internal node that node's
2637  * parent must be adjusted to the element's new location.
2638  *
2639  * XXX deadlock potential here with our exclusive locks
2640  */
2641 int
2642 btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2643                  hammer_btree_elm_t elm)
2644 {
2645         hammer_node_t child;
2646         int error;
2647
2648         error = 0;
2649
2650         switch(elm->base.btype) {
2651         case HAMMER_BTREE_TYPE_INTERNAL:
2652         case HAMMER_BTREE_TYPE_LEAF:
2653                 child = hammer_get_node(trans, elm->internal.subtree_offset,
2654                                         0, &error);
2655                 if (error == 0) {
2656                         hammer_modify_node_field(trans, child, parent);
2657                         child->ondisk->parent = node->node_offset;
2658                         hammer_modify_node_done(child);
2659                         hammer_rel_node(child);
2660                 }
2661                 break;
2662         default:
2663                 break;
2664         }
2665         return(error);
2666 }
2667
2668 /*
2669  * Initialize the root of a recursive B-Tree node lock list structure.
2670  */
2671 void
2672 hammer_node_lock_init(hammer_node_lock_t parent, hammer_node_t node)
2673 {
2674         TAILQ_INIT(&parent->list);
2675         parent->parent = NULL;
2676         parent->node = node;
2677         parent->index = -1;
2678         parent->count = node->ondisk->count;
2679         parent->copy = NULL;
2680         parent->flags = 0;
2681 }
2682
2683 /*
2684  * Initialize a cache of hammer_node_lock's including space allocated
2685  * for node copies.
2686  *
2687  * This is used by the rebalancing code to preallocate the copy space
2688  * for ~4096 B-Tree nodes (16MB of data) prior to acquiring any HAMMER
2689  * locks, otherwise we can blow out the pageout daemon's emergency
2690  * reserve and deadlock it.
2691  *
2692  * NOTE: HAMMER_NODE_LOCK_LCACHE is not set on items cached in the lcache.
2693  *       The flag is set when the item is pulled off the cache for use.
2694  */
2695 void
2696 hammer_btree_lcache_init(hammer_mount_t hmp, hammer_node_lock_t lcache,
2697                          int depth)
2698 {
2699         hammer_node_lock_t item;
2700         int count;
2701
2702         for (count = 1; depth; --depth)
2703                 count *= HAMMER_BTREE_LEAF_ELMS;
2704         bzero(lcache, sizeof(*lcache));
2705         TAILQ_INIT(&lcache->list);
2706         while (count) {
2707                 item = kmalloc(sizeof(*item), hmp->m_misc, M_WAITOK|M_ZERO);
2708                 item->copy = kmalloc(sizeof(*item->copy),
2709                                      hmp->m_misc, M_WAITOK);
2710                 TAILQ_INIT(&item->list);
2711                 TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2712                 --count;
2713         }
2714 }
2715
2716 void
2717 hammer_btree_lcache_free(hammer_mount_t hmp, hammer_node_lock_t lcache)
2718 {
2719         hammer_node_lock_t item;
2720
2721         while ((item = TAILQ_FIRST(&lcache->list)) != NULL) {
2722                 TAILQ_REMOVE(&lcache->list, item, entry);
2723                 KKASSERT(item->copy);
2724                 KKASSERT(TAILQ_EMPTY(&item->list));
2725                 kfree(item->copy, hmp->m_misc);
2726                 kfree(item, hmp->m_misc);
2727         }
2728         KKASSERT(lcache->copy == NULL);
2729 }
2730
2731 /*
2732  * Exclusively lock all the children of node.  This is used by the split
2733  * code to prevent anyone from accessing the children of a cursor node
2734  * while we fix-up its parent offset.
2735  *
2736  * If we don't lock the children we can really mess up cursors which block
2737  * trying to cursor-up into our node.
2738  *
2739  * On failure EDEADLK (or some other error) is returned.  If a deadlock
2740  * error is returned the cursor is adjusted to block on termination.
2741  *
2742  * The caller is responsible for managing parent->node, the root's node
2743  * is usually aliased from a cursor.
2744  */
2745 int
2746 hammer_btree_lock_children(hammer_cursor_t cursor, int depth,
2747                            hammer_node_lock_t parent,
2748                            hammer_node_lock_t lcache)
2749 {
2750         hammer_node_t node;
2751         hammer_node_lock_t item;
2752         hammer_node_ondisk_t ondisk;
2753         hammer_btree_elm_t elm;
2754         hammer_node_t child;
2755         struct hammer_mount *hmp;
2756         int error;
2757         int i;
2758
2759         node = parent->node;
2760         ondisk = node->ondisk;
2761         error = 0;
2762         hmp = cursor->trans->hmp;
2763
2764         /*
2765          * We really do not want to block on I/O with exclusive locks held,
2766          * pre-get the children before trying to lock the mess.  This is
2767          * only done one-level deep for now.
2768          */
2769         for (i = 0; i < ondisk->count; ++i) {
2770                 ++hammer_stats_btree_elements;
2771                 elm = &ondisk->elms[i];
2772                 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2773                     elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2774                         continue;
2775                 }
2776                 child = hammer_get_node(cursor->trans,
2777                                         elm->internal.subtree_offset,
2778                                         0, &error);
2779                 if (child)
2780                         hammer_rel_node(child);
2781         }
2782
2783         /*
2784          * Do it for real
2785          */
2786         for (i = 0; error == 0 && i < ondisk->count; ++i) {
2787                 ++hammer_stats_btree_elements;
2788                 elm = &ondisk->elms[i];
2789
2790                 switch(elm->base.btype) {
2791                 case HAMMER_BTREE_TYPE_INTERNAL:
2792                 case HAMMER_BTREE_TYPE_LEAF:
2793                         KKASSERT(elm->internal.subtree_offset != 0);
2794                         child = hammer_get_node(cursor->trans,
2795                                                 elm->internal.subtree_offset,
2796                                                 0, &error);
2797                         break;
2798                 default:
2799                         child = NULL;
2800                         break;
2801                 }
2802                 if (child) {
2803                         if (hammer_lock_ex_try(&child->lock) != 0) {
2804                                 if (cursor->deadlk_node == NULL) {
2805                                         cursor->deadlk_node = child;
2806                                         hammer_ref_node(cursor->deadlk_node);
2807                                 }
2808                                 error = EDEADLK;
2809                                 hammer_rel_node(child);
2810                         } else {
2811                                 if (lcache) {
2812                                         item = TAILQ_FIRST(&lcache->list);
2813                                         KKASSERT(item != NULL);
2814                                         item->flags |= HAMMER_NODE_LOCK_LCACHE;
2815                                         TAILQ_REMOVE(&lcache->list,
2816                                                      item, entry);
2817                                 } else {
2818                                         item = kmalloc(sizeof(*item),
2819                                                        hmp->m_misc,
2820                                                        M_WAITOK|M_ZERO);
2821                                         TAILQ_INIT(&item->list);
2822                                 }
2823
2824                                 TAILQ_INSERT_TAIL(&parent->list, item, entry);
2825                                 item->parent = parent;
2826                                 item->node = child;
2827                                 item->index = i;
2828                                 item->count = child->ondisk->count;
2829
2830                                 /*
2831                                  * Recurse (used by the rebalancing code)
2832                                  */
2833                                 if (depth > 1 && elm->base.btype == HAMMER_BTREE_TYPE_INTERNAL) {
2834                                         error = hammer_btree_lock_children(
2835                                                         cursor,
2836                                                         depth - 1,
2837                                                         item,
2838                                                         lcache);
2839                                 }
2840                         }
2841                 }
2842         }
2843         if (error)
2844                 hammer_btree_unlock_children(hmp, parent, lcache);
2845         return(error);
2846 }
2847
2848 /*
2849  * Create an in-memory copy of all B-Tree nodes listed, recursively,
2850  * including the parent.
2851  */
2852 void
2853 hammer_btree_lock_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2854 {
2855         hammer_mount_t hmp = cursor->trans->hmp;
2856         hammer_node_lock_t item;
2857
2858         if (parent->copy == NULL) {
2859                 KKASSERT((parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0);
2860                 parent->copy = kmalloc(sizeof(*parent->copy),
2861                                        hmp->m_misc, M_WAITOK);
2862         }
2863         KKASSERT((parent->flags & HAMMER_NODE_LOCK_UPDATED) == 0);
2864         *parent->copy = *parent->node->ondisk;
2865         TAILQ_FOREACH(item, &parent->list, entry) {
2866                 hammer_btree_lock_copy(cursor, item);
2867         }
2868 }
2869
2870 /*
2871  * Recursively sync modified copies to the media.
2872  */
2873 int
2874 hammer_btree_sync_copy(hammer_cursor_t cursor, hammer_node_lock_t parent)
2875 {
2876         hammer_node_lock_t item;
2877         int count = 0;
2878
2879         if (parent->flags & HAMMER_NODE_LOCK_UPDATED) {
2880                 ++count;
2881                 hammer_modify_node_all(cursor->trans, parent->node);
2882                 *parent->node->ondisk = *parent->copy;
2883                 hammer_modify_node_done(parent->node);
2884                 if (parent->copy->type == HAMMER_BTREE_TYPE_DELETED) {
2885                         hammer_flush_node(parent->node, 0);
2886                         hammer_delete_node(cursor->trans, parent->node);
2887                 }
2888         }
2889         TAILQ_FOREACH(item, &parent->list, entry) {
2890                 count += hammer_btree_sync_copy(cursor, item);
2891         }
2892         return(count);
2893 }
2894
2895 /*
2896  * Release previously obtained node locks.  The caller is responsible for
2897  * cleaning up parent->node itself (its usually just aliased from a cursor),
2898  * but this function will take care of the copies.
2899  *
2900  * NOTE: The root node is not placed in the lcache and node->copy is not
2901  *       deallocated when lcache != NULL.
2902  */
2903 void
2904 hammer_btree_unlock_children(hammer_mount_t hmp, hammer_node_lock_t parent,
2905                              hammer_node_lock_t lcache)
2906 {
2907         hammer_node_lock_t item;
2908         hammer_node_ondisk_t copy;
2909
2910         while ((item = TAILQ_FIRST(&parent->list)) != NULL) {
2911                 TAILQ_REMOVE(&parent->list, item, entry);
2912                 hammer_btree_unlock_children(hmp, item, lcache);
2913                 hammer_unlock(&item->node->lock);
2914                 hammer_rel_node(item->node);
2915                 if (lcache) {
2916                         /*
2917                          * NOTE: When placing the item back in the lcache
2918                          *       the flag is cleared by the bzero().
2919                          *       Remaining fields are cleared as a safety
2920                          *       measure.
2921                          */
2922                         KKASSERT(item->flags & HAMMER_NODE_LOCK_LCACHE);
2923                         KKASSERT(TAILQ_EMPTY(&item->list));
2924                         copy = item->copy;
2925                         bzero(item, sizeof(*item));
2926                         TAILQ_INIT(&item->list);
2927                         item->copy = copy;
2928                         if (copy)
2929                                 bzero(copy, sizeof(*copy));
2930                         TAILQ_INSERT_TAIL(&lcache->list, item, entry);
2931                 } else {
2932                         kfree(item, hmp->m_misc);
2933                 }
2934         }
2935         if (parent->copy && (parent->flags & HAMMER_NODE_LOCK_LCACHE) == 0) {
2936                 kfree(parent->copy, hmp->m_misc);
2937                 parent->copy = NULL;    /* safety */
2938         }
2939 }
2940
2941 /************************************************************************
2942  *                         MISCELLANIOUS SUPPORT                        *
2943  ************************************************************************/
2944
2945 /*
2946  * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
2947  *
2948  * Note that for this particular function a return value of -1, 0, or +1
2949  * can denote a match if create_tid is otherwise discounted.  A create_tid
2950  * of zero is considered to be 'infinity' in comparisons.
2951  *
2952  * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
2953  */
2954 int
2955 hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2956 {
2957         if (key1->localization < key2->localization)
2958                 return(-5);
2959         if (key1->localization > key2->localization)
2960                 return(5);
2961
2962         if (key1->obj_id < key2->obj_id)
2963                 return(-4);
2964         if (key1->obj_id > key2->obj_id)
2965                 return(4);
2966
2967         if (key1->rec_type < key2->rec_type)
2968                 return(-3);
2969         if (key1->rec_type > key2->rec_type)
2970                 return(3);
2971
2972         if (key1->key < key2->key)
2973                 return(-2);
2974         if (key1->key > key2->key)
2975                 return(2);
2976
2977         /*
2978          * A create_tid of zero indicates a record which is undeletable
2979          * and must be considered to have a value of positive infinity.
2980          */
2981         if (key1->create_tid == 0) {
2982                 if (key2->create_tid == 0)
2983                         return(0);
2984                 return(1);
2985         }
2986         if (key2->create_tid == 0)
2987                 return(-1);
2988         if (key1->create_tid < key2->create_tid)
2989                 return(-1);
2990         if (key1->create_tid > key2->create_tid)
2991                 return(1);
2992         return(0);
2993 }
2994
2995 /*
2996  * Test a timestamp against an element to determine whether the
2997  * element is visible.  A timestamp of 0 means 'infinity'.
2998  */
2999 int
3000 hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
3001 {
3002         if (asof == 0) {
3003                 if (base->delete_tid)
3004                         return(1);
3005                 return(0);
3006         }
3007         if (asof < base->create_tid)
3008                 return(-1);
3009         if (base->delete_tid && asof >= base->delete_tid)
3010                 return(1);
3011         return(0);
3012 }
3013
3014 /*
3015  * Create a separator half way inbetween key1 and key2.  For fields just
3016  * one unit apart, the separator will match key2.  key1 is on the left-hand
3017  * side and key2 is on the right-hand side.
3018  *
3019  * key2 must be >= the separator.  It is ok for the separator to match key2.
3020  *
3021  * NOTE: Even if key1 does not match key2, the separator may wind up matching
3022  * key2.
3023  *
3024  * NOTE: It might be beneficial to just scrap this whole mess and just
3025  * set the separator to key2.
3026  */
3027 #define MAKE_SEPARATOR(key1, key2, dest, field) \
3028         dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
3029
3030 static void
3031 hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
3032                       hammer_base_elm_t dest)
3033 {
3034         bzero(dest, sizeof(*dest));
3035
3036         dest->rec_type = key2->rec_type;
3037         dest->key = key2->key;
3038         dest->obj_id = key2->obj_id;
3039         dest->create_tid = key2->create_tid;
3040
3041         MAKE_SEPARATOR(key1, key2, dest, localization);
3042         if (key1->localization == key2->localization) {
3043                 MAKE_SEPARATOR(key1, key2, dest, obj_id);
3044                 if (key1->obj_id == key2->obj_id) {
3045                         MAKE_SEPARATOR(key1, key2, dest, rec_type);
3046                         if (key1->rec_type == key2->rec_type) {
3047                                 MAKE_SEPARATOR(key1, key2, dest, key);
3048                                 /*
3049                                  * Don't bother creating a separator for
3050                                  * create_tid, which also conveniently avoids
3051                                  * having to handle the create_tid == 0
3052                                  * (infinity) case.  Just leave create_tid
3053                                  * set to key2.
3054                                  *
3055                                  * Worst case, dest matches key2 exactly,
3056                                  * which is acceptable.
3057                                  */
3058                         }
3059                 }
3060         }
3061 }
3062
3063 #undef MAKE_SEPARATOR
3064
3065 /*
3066  * Return whether a generic internal or leaf node is full
3067  */
3068 static int
3069 btree_node_is_full(hammer_node_ondisk_t node)
3070 {
3071         switch(node->type) {
3072         case HAMMER_BTREE_TYPE_INTERNAL:
3073                 if (node->count == HAMMER_BTREE_INT_ELMS)
3074                         return(1);
3075                 break;
3076         case HAMMER_BTREE_TYPE_LEAF:
3077                 if (node->count == HAMMER_BTREE_LEAF_ELMS)
3078                         return(1);
3079                 break;
3080         default:
3081                 panic("illegal btree subtype");
3082         }
3083         return(0);
3084 }
3085
3086 #if 0
3087 static int
3088 btree_max_elements(u_int8_t type)
3089 {
3090         if (type == HAMMER_BTREE_TYPE_LEAF)
3091                 return(HAMMER_BTREE_LEAF_ELMS);
3092         if (type == HAMMER_BTREE_TYPE_INTERNAL)
3093                 return(HAMMER_BTREE_INT_ELMS);
3094         panic("btree_max_elements: bad type %d\n", type);
3095 }
3096 #endif
3097
3098 void
3099 hammer_print_btree_node(hammer_node_ondisk_t ondisk)
3100 {
3101         hammer_btree_elm_t elm;
3102         int i;
3103
3104         kprintf("node %p count=%d parent=%016llx type=%c\n",
3105                 ondisk, ondisk->count,
3106                 (long long)ondisk->parent, ondisk->type);
3107
3108         /*
3109          * Dump both boundary elements if an internal node
3110          */
3111         if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
3112                 for (i = 0; i <= ondisk->count; ++i) {
3113                         elm = &ondisk->elms[i];
3114                         hammer_print_btree_elm(elm, ondisk->type, i);
3115                 }
3116         } else {
3117                 for (i = 0; i < ondisk->count; ++i) {
3118                         elm = &ondisk->elms[i];
3119                         hammer_print_btree_elm(elm, ondisk->type, i);
3120                 }
3121         }
3122 }
3123
3124 void
3125 hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
3126 {
3127         kprintf("  %2d", i);
3128         kprintf("\tobj_id       = %016llx\n", (long long)elm->base.obj_id);
3129         kprintf("\tkey          = %016llx\n", (long long)elm->base.key);
3130         kprintf("\tcreate_tid   = %016llx\n", (long long)elm->base.create_tid);
3131         kprintf("\tdelete_tid   = %016llx\n", (long long)elm->base.delete_tid);
3132         kprintf("\trec_type     = %04x\n", elm->base.rec_type);
3133         kprintf("\tobj_type     = %02x\n", elm->base.obj_type);
3134         kprintf("\tbtype        = %02x (%c)\n",
3135                 elm->base.btype,
3136                 (elm->base.btype ? elm->base.btype : '?'));
3137         kprintf("\tlocalization = %02x\n", elm->base.localization);
3138
3139         switch(type) {
3140         case HAMMER_BTREE_TYPE_INTERNAL:
3141                 kprintf("\tsubtree_off  = %016llx\n",
3142                         (long long)elm->internal.subtree_offset);
3143                 break;
3144         case HAMMER_BTREE_TYPE_RECORD:
3145                 kprintf("\tdata_offset  = %016llx\n",
3146                         (long long)elm->leaf.data_offset);
3147                 kprintf("\tdata_len     = %08x\n", elm->leaf.data_len);
3148                 kprintf("\tdata_crc     = %08x\n", elm->leaf.data_crc);
3149                 break;
3150         }
3151 }