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