Ensure AP is associated before we return.
[dragonfly.git] / sys / vfs / hammer / hammer_btree.c
CommitLineData
427e5fc6 1/*
b84de5af 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved.
427e5fc6
MD
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 *
b3bad96f 34 * $DragonFly: src/sys/vfs/hammer/hammer_btree.c,v 1.63 2008/07/05 18:59:27 dillon Exp $
427e5fc6
MD
35 */
36
37/*
8cd0a023 38 * HAMMER B-Tree index
427e5fc6
MD
39 *
40 * HAMMER implements a modified B+Tree. In documentation this will
9944ae54 41 * simply be refered to as the HAMMER B-Tree. Basically a HAMMER B-Tree
427e5fc6
MD
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
8cd0a023 45 * otherwords, we have boundary elements at the two ends of a B-Tree node
427e5fc6
MD
46 * instead of sub-tree pointers.
47 *
8cd0a023 48 * A B-Tree internal node looks like this:
427e5fc6
MD
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 *
8cd0a023 53 * A B-Tree leaf node basically looks like this:
427e5fc6
MD
54 *
55 * L L L L L L L L <-- leaf elemenets
56 *
8cd0a023
MD
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.
427e5fc6 59 *
8cd0a023
MD
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.
427e5fc6 67 *
8cd0a023
MD
68 * B-Trees also make the stacking of trees fairly straightforward.
69 *
fe7678ee
MD
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
f36a9737
MD
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).
8cd0a023
MD
80 */
81#include "hammer.h"
82#include <sys/buf.h>
83#include <sys/buf2.h>
66325755 84
8cd0a023
MD
85static int btree_search(hammer_cursor_t cursor, int flags);
86static int btree_split_internal(hammer_cursor_t cursor);
87static int btree_split_leaf(hammer_cursor_t cursor);
46fe7ae1 88static int btree_remove(hammer_cursor_t cursor);
fe7678ee 89static int btree_node_is_full(hammer_node_ondisk_t node);
602c6cb8
MD
90static int hammer_btree_mirror_propagate(hammer_transaction_t trans,
91 hammer_node_t node, int index, hammer_tid_t mirror_tid);
8cd0a023
MD
92static void hammer_make_separator(hammer_base_elm_t key1,
93 hammer_base_elm_t key2, hammer_base_elm_t dest);
66325755
MD
94
95/*
8cd0a023
MD
96 * Iterate records after a search. The cursor is iterated forwards past
97 * the current record until a record matching the key-range requirements
98 * is found. ENOENT is returned if the iteration goes past the ending
6a37e7e4 99 * key.
66325755 100 *
d26d0ae9
MD
101 * The iteration is inclusive of key_beg and can be inclusive or exclusive
102 * of key_end depending on whether HAMMER_CURSOR_END_INCLUSIVE is set.
66325755 103 *
eaeff70d 104 * When doing an as-of search (cursor->asof != 0), key_beg.create_tid
9582c7da 105 * may be modified by B-Tree functions.
d5530d22 106 *
8cd0a023 107 * cursor->key_beg may or may not be modified by this function during
d26d0ae9
MD
108 * the iteration. XXX future - in case of an inverted lock we may have
109 * to reinitiate the lookup and set key_beg to properly pick up where we
110 * left off.
6a37e7e4
MD
111 *
112 * NOTE! EDEADLK *CANNOT* be returned by this procedure.
66325755
MD
113 */
114int
8cd0a023 115hammer_btree_iterate(hammer_cursor_t cursor)
66325755 116{
8cd0a023
MD
117 hammer_node_ondisk_t node;
118 hammer_btree_elm_t elm;
66325755
MD
119 int error;
120 int r;
121 int s;
122
123 /*
8cd0a023 124 * Skip past the current record
66325755 125 */
8cd0a023 126 node = cursor->node->ondisk;
a89aec1b
MD
127 if (node == NULL)
128 return(ENOENT);
c0ade690
MD
129 if (cursor->index < node->count &&
130 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
66325755 131 ++cursor->index;
c0ade690 132 }
66325755 133
8cd0a023
MD
134 /*
135 * Loop until an element is found or we are done.
136 */
66325755
MD
137 for (;;) {
138 /*
8cd0a023
MD
139 * We iterate up the tree and then index over one element
140 * while we are at the last element in the current node.
141 *
47197d71 142 * If we are at the root of the filesystem, cursor_up
8cd0a023
MD
143 * returns ENOENT.
144 *
66325755
MD
145 * XXX this could be optimized by storing the information in
146 * the parent reference.
195c19a1
MD
147 *
148 * XXX we can lose the node lock temporarily, this could mess
149 * up our scan.
66325755 150 */
47637bff 151 ++hammer_stats_btree_iterations;
77fec802
MD
152 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
153
8cd0a023 154 if (cursor->index == node->count) {
a84a197d
MD
155 if (hammer_debug_btree) {
156 kprintf("BRACKETU %016llx[%d] -> %016llx[%d] (td=%p)\n",
157 cursor->node->node_offset,
158 cursor->index,
159 (cursor->parent ? cursor->parent->node_offset : -1),
160 cursor->parent_index,
161 curthread);
162 }
163 KKASSERT(cursor->parent == NULL || cursor->parent->ondisk->elms[cursor->parent_index].internal.subtree_offset == cursor->node->node_offset);
6a37e7e4 164 error = hammer_cursor_up(cursor);
8cd0a023
MD
165 if (error)
166 break;
46fe7ae1 167 /* reload stale pointer */
8cd0a023
MD
168 node = cursor->node->ondisk;
169 KKASSERT(cursor->index != node->count);
2f85fa4d
MD
170
171 /*
172 * If we are reblocking we want to return internal
173 * nodes.
174 */
175 if (cursor->flags & HAMMER_CURSOR_REBLOCKING) {
176 cursor->flags |= HAMMER_CURSOR_ATEDISK;
177 return(0);
178 }
8cd0a023
MD
179 ++cursor->index;
180 continue;
66325755
MD
181 }
182
183 /*
d26d0ae9
MD
184 * Check internal or leaf element. Determine if the record
185 * at the cursor has gone beyond the end of our range.
66325755 186 *
47197d71 187 * We recurse down through internal nodes.
66325755 188 */
8cd0a023
MD
189 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
190 elm = &node->elms[cursor->index];
c82af904 191
d26d0ae9
MD
192 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
193 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
b3deaf57 194 if (hammer_debug_btree) {
2f85fa4d 195 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d (td=%p)\n",
eaeff70d
MD
196 cursor->node->node_offset,
197 cursor->index,
b3deaf57
MD
198 elm[0].internal.base.obj_id,
199 elm[0].internal.base.rec_type,
200 elm[0].internal.base.key,
2f85fa4d 201 elm[0].internal.base.localization,
a84a197d
MD
202 r,
203 curthread
b3deaf57 204 );
2f85fa4d 205 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
eaeff70d
MD
206 cursor->node->node_offset,
207 cursor->index + 1,
b3deaf57
MD
208 elm[1].internal.base.obj_id,
209 elm[1].internal.base.rec_type,
210 elm[1].internal.base.key,
2f85fa4d 211 elm[1].internal.base.localization,
b3deaf57
MD
212 s
213 );
214 }
215
d26d0ae9
MD
216 if (r < 0) {
217 error = ENOENT;
218 break;
66325755 219 }
fe7678ee
MD
220 if (r == 0 && (cursor->flags &
221 HAMMER_CURSOR_END_INCLUSIVE) == 0) {
d26d0ae9 222 error = ENOENT;
8cd0a023 223 break;
d26d0ae9
MD
224 }
225 KKASSERT(s <= 0);
6a37e7e4
MD
226
227 /*
f36a9737 228 * Better not be zero
6a37e7e4 229 */
f36a9737
MD
230 KKASSERT(elm->internal.subtree_offset != 0);
231
c82af904
MD
232 /*
233 * If running the mirror filter see if we can skip
234 * the entire sub-tree.
235 */
236 if (cursor->flags & HAMMER_CURSOR_MIRROR_FILTERED) {
237 if (elm->internal.mirror_tid <
238 cursor->mirror_tid) {
239 ++cursor->index;
240 continue;
241 }
242 }
243
f36a9737
MD
244 error = hammer_cursor_down(cursor);
245 if (error)
246 break;
247 KKASSERT(cursor->index == 0);
46fe7ae1
MD
248 /* reload stale pointer */
249 node = cursor->node->ondisk;
fe7678ee 250 continue;
d26d0ae9
MD
251 } else {
252 elm = &node->elms[cursor->index];
253 r = hammer_btree_cmp(&cursor->key_end, &elm->base);
b3deaf57 254 if (hammer_debug_btree) {
2f85fa4d 255 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
eaeff70d
MD
256 cursor->node->node_offset,
257 cursor->index,
258 (elm[0].leaf.base.btype ?
259 elm[0].leaf.base.btype : '?'),
b3deaf57
MD
260 elm[0].leaf.base.obj_id,
261 elm[0].leaf.base.rec_type,
262 elm[0].leaf.base.key,
2f85fa4d 263 elm[0].leaf.base.localization,
b3deaf57
MD
264 r
265 );
266 }
d26d0ae9
MD
267 if (r < 0) {
268 error = ENOENT;
269 break;
270 }
b33e2cc0
MD
271
272 /*
273 * We support both end-inclusive and
274 * end-exclusive searches.
275 */
276 if (r == 0 &&
277 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
278 error = ENOENT;
279 break;
280 }
281
fe7678ee
MD
282 switch(elm->leaf.base.btype) {
283 case HAMMER_BTREE_TYPE_RECORD:
284 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
285 hammer_btree_chkts(cursor->asof, &elm->base)) {
286 ++cursor->index;
287 continue;
288 }
289 break;
fe7678ee
MD
290 default:
291 error = EINVAL;
292 break;
d26d0ae9 293 }
fe7678ee
MD
294 if (error)
295 break;
66325755 296 }
46fe7ae1
MD
297 /*
298 * node pointer invalid after loop
299 */
66325755
MD
300
301 /*
d26d0ae9 302 * Return entry
66325755 303 */
b3deaf57
MD
304 if (hammer_debug_btree) {
305 int i = cursor->index;
306 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
2f85fa4d 307 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
b3deaf57
MD
308 cursor->node, i,
309 elm->internal.base.obj_id,
310 elm->internal.base.rec_type,
2f85fa4d
MD
311 elm->internal.base.key,
312 elm->internal.base.localization
b3deaf57
MD
313 );
314 }
d26d0ae9 315 return(0);
427e5fc6 316 }
66325755 317 return(error);
427e5fc6
MD
318}
319
32c90105
MD
320/*
321 * Iterate in the reverse direction. This is used by the pruning code to
322 * avoid overlapping records.
323 */
324int
325hammer_btree_iterate_reverse(hammer_cursor_t cursor)
326{
327 hammer_node_ondisk_t node;
328 hammer_btree_elm_t elm;
329 int error;
330 int r;
331 int s;
332
333 /*
334 * Skip past the current record. For various reasons the cursor
335 * may end up set to -1 or set to point at the end of the current
336 * node. These cases must be addressed.
337 */
338 node = cursor->node->ondisk;
339 if (node == NULL)
340 return(ENOENT);
341 if (cursor->index != -1 &&
342 (cursor->flags & HAMMER_CURSOR_ATEDISK)) {
343 --cursor->index;
344 }
345 if (cursor->index == cursor->node->ondisk->count)
346 --cursor->index;
347
348 /*
349 * Loop until an element is found or we are done.
350 */
351 for (;;) {
77fec802
MD
352 ++hammer_stats_btree_iterations;
353 hammer_flusher_clean_loose_ios(cursor->trans->hmp);
354
32c90105
MD
355 /*
356 * We iterate up the tree and then index over one element
357 * while we are at the last element in the current node.
32c90105
MD
358 */
359 if (cursor->index == -1) {
360 error = hammer_cursor_up(cursor);
361 if (error) {
362 cursor->index = 0; /* sanity */
363 break;
364 }
365 /* reload stale pointer */
366 node = cursor->node->ondisk;
367 KKASSERT(cursor->index != node->count);
368 --cursor->index;
369 continue;
370 }
371
372 /*
373 * Check internal or leaf element. Determine if the record
374 * at the cursor has gone beyond the end of our range.
375 *
47197d71 376 * We recurse down through internal nodes.
32c90105
MD
377 */
378 KKASSERT(cursor->index != node->count);
379 if (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
380 elm = &node->elms[cursor->index];
381 r = hammer_btree_cmp(&cursor->key_end, &elm[0].base);
382 s = hammer_btree_cmp(&cursor->key_beg, &elm[1].base);
383 if (hammer_debug_btree) {
2f85fa4d 384 kprintf("BRACKETL %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
32c90105
MD
385 cursor->node->node_offset,
386 cursor->index,
387 elm[0].internal.base.obj_id,
388 elm[0].internal.base.rec_type,
389 elm[0].internal.base.key,
2f85fa4d 390 elm[0].internal.base.localization,
32c90105
MD
391 r
392 );
2f85fa4d 393 kprintf("BRACKETR %016llx[%d] %016llx %02x %016llx lo=%02x %d\n",
32c90105
MD
394 cursor->node->node_offset,
395 cursor->index + 1,
396 elm[1].internal.base.obj_id,
397 elm[1].internal.base.rec_type,
398 elm[1].internal.base.key,
2f85fa4d 399 elm[1].internal.base.localization,
32c90105
MD
400 s
401 );
402 }
403
404 if (s >= 0) {
405 error = ENOENT;
406 break;
407 }
408 KKASSERT(r >= 0);
409
410 /*
f36a9737 411 * Better not be zero
32c90105 412 */
f36a9737
MD
413 KKASSERT(elm->internal.subtree_offset != 0);
414
415 error = hammer_cursor_down(cursor);
416 if (error)
417 break;
418 KKASSERT(cursor->index == 0);
32c90105
MD
419 /* reload stale pointer */
420 node = cursor->node->ondisk;
f36a9737
MD
421
422 /* this can assign -1 if the leaf was empty */
423 cursor->index = node->count - 1;
32c90105
MD
424 continue;
425 } else {
426 elm = &node->elms[cursor->index];
427 s = hammer_btree_cmp(&cursor->key_beg, &elm->base);
428 if (hammer_debug_btree) {
2f85fa4d 429 kprintf("ELEMENT %016llx:%d %c %016llx %02x %016llx lo=%02x %d\n",
32c90105
MD
430 cursor->node->node_offset,
431 cursor->index,
432 (elm[0].leaf.base.btype ?
433 elm[0].leaf.base.btype : '?'),
434 elm[0].leaf.base.obj_id,
435 elm[0].leaf.base.rec_type,
436 elm[0].leaf.base.key,
2f85fa4d 437 elm[0].leaf.base.localization,
32c90105
MD
438 s
439 );
440 }
441 if (s > 0) {
442 error = ENOENT;
443 break;
444 }
445
446 switch(elm->leaf.base.btype) {
447 case HAMMER_BTREE_TYPE_RECORD:
448 if ((cursor->flags & HAMMER_CURSOR_ASOF) &&
449 hammer_btree_chkts(cursor->asof, &elm->base)) {
450 --cursor->index;
451 continue;
452 }
453 break;
32c90105
MD
454 default:
455 error = EINVAL;
456 break;
457 }
458 if (error)
459 break;
460 }
461 /*
462 * node pointer invalid after loop
463 */
464
465 /*
466 * Return entry
467 */
468 if (hammer_debug_btree) {
469 int i = cursor->index;
470 hammer_btree_elm_t elm = &cursor->node->ondisk->elms[i];
2f85fa4d 471 kprintf("ITERATE %p:%d %016llx %02x %016llx lo=%02x\n",
32c90105
MD
472 cursor->node, i,
473 elm->internal.base.obj_id,
474 elm->internal.base.rec_type,
2f85fa4d
MD
475 elm->internal.base.key,
476 elm->internal.base.localization
32c90105
MD
477 );
478 }
479 return(0);
480 }
481 return(error);
482}
483
427e5fc6 484/*
8cd0a023 485 * Lookup cursor->key_beg. 0 is returned on success, ENOENT if the entry
6a37e7e4
MD
486 * could not be found, EDEADLK if inserting and a retry is needed, and a
487 * fatal error otherwise. When retrying, the caller must terminate the
eaeff70d 488 * cursor and reinitialize it. EDEADLK cannot be returned if not inserting.
8cd0a023
MD
489 *
490 * The cursor is suitably positioned for a deletion on success, and suitably
eaeff70d
MD
491 * positioned for an insertion on ENOENT if HAMMER_CURSOR_INSERT was
492 * specified.
427e5fc6 493 *
47197d71 494 * The cursor may begin anywhere, the search will traverse the tree in
8cd0a023 495 * either direction to locate the requested element.
eaeff70d
MD
496 *
497 * Most of the logic implementing historical searches is handled here. We
9582c7da
MD
498 * do an initial lookup with create_tid set to the asof TID. Due to the
499 * way records are laid out, a backwards iteration may be required if
eaeff70d
MD
500 * ENOENT is returned to locate the historical record. Here's the
501 * problem:
502 *
9582c7da 503 * create_tid: 10 15 20
eaeff70d
MD
504 * LEAF1 LEAF2
505 * records: (11) (18)
506 *
9582c7da
MD
507 * Lets say we want to do a lookup AS-OF timestamp 17. We will traverse
508 * LEAF2 but the only record in LEAF2 has a create_tid of 18, which is
509 * not visible and thus causes ENOENT to be returned. We really need
510 * to check record 11 in LEAF1. If it also fails then the search fails
511 * (e.g. it might represent the range 11-16 and thus still not match our
f36a9737
MD
512 * AS-OF timestamp of 17). Note that LEAF1 could be empty, requiring
513 * further iterations.
b33e2cc0 514 *
9582c7da
MD
515 * If this case occurs btree_search() will set HAMMER_CURSOR_CREATE_CHECK
516 * and the cursor->create_check TID if an iteration might be needed.
517 * In the above example create_check would be set to 14.
427e5fc6
MD
518 */
519int
8cd0a023 520hammer_btree_lookup(hammer_cursor_t cursor)
427e5fc6 521{
66325755
MD
522 int error;
523
cb51be26 524 ++hammer_stats_btree_lookups;
d5530d22 525 if (cursor->flags & HAMMER_CURSOR_ASOF) {
eaeff70d 526 KKASSERT((cursor->flags & HAMMER_CURSOR_INSERT) == 0);
9582c7da 527 cursor->key_beg.create_tid = cursor->asof;
eaeff70d 528 for (;;) {
9582c7da 529 cursor->flags &= ~HAMMER_CURSOR_CREATE_CHECK;
d5530d22 530 error = btree_search(cursor, 0);
b33e2cc0 531 if (error != ENOENT ||
9582c7da 532 (cursor->flags & HAMMER_CURSOR_CREATE_CHECK) == 0) {
b33e2cc0
MD
533 /*
534 * Stop if no error.
535 * Stop if error other then ENOENT.
536 * Stop if ENOENT and not special case.
537 */
eaeff70d
MD
538 break;
539 }
32c90105
MD
540 if (hammer_debug_btree) {
541 kprintf("CREATE_CHECK %016llx\n",
542 cursor->create_check);
543 }
9582c7da 544 cursor->key_beg.create_tid = cursor->create_check;
eaeff70d
MD
545 /* loop */
546 }
d5530d22
MD
547 } else {
548 error = btree_search(cursor, 0);
549 }
bf3b416b 550 if (error == 0)
8cd0a023 551 error = hammer_btree_extract(cursor, cursor->flags);
66325755
MD
552 return(error);
553}
554
d26d0ae9
MD
555/*
556 * Execute the logic required to start an iteration. The first record
557 * located within the specified range is returned and iteration control
558 * flags are adjusted for successive hammer_btree_iterate() calls.
559 */
560int
561hammer_btree_first(hammer_cursor_t cursor)
562{
563 int error;
564
565 error = hammer_btree_lookup(cursor);
566 if (error == ENOENT) {
567 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
568 error = hammer_btree_iterate(cursor);
569 }
570 cursor->flags |= HAMMER_CURSOR_ATEDISK;
571 return(error);
572}
573
32c90105
MD
574/*
575 * Similarly but for an iteration in the reverse direction.
814387f6
MD
576 *
577 * Set ATEDISK when iterating backwards to skip the current entry,
578 * which after an ENOENT lookup will be pointing beyond our end point.
32c90105
MD
579 */
580int
581hammer_btree_last(hammer_cursor_t cursor)
582{
583 struct hammer_base_elm save;
584 int error;
585
586 save = cursor->key_beg;
587 cursor->key_beg = cursor->key_end;
588 error = hammer_btree_lookup(cursor);
589 cursor->key_beg = save;
590 if (error == ENOENT ||
591 (cursor->flags & HAMMER_CURSOR_END_INCLUSIVE) == 0) {
814387f6 592 cursor->flags |= HAMMER_CURSOR_ATEDISK;
32c90105
MD
593 error = hammer_btree_iterate_reverse(cursor);
594 }
595 cursor->flags |= HAMMER_CURSOR_ATEDISK;
596 return(error);
597}
598
8cd0a023
MD
599/*
600 * Extract the record and/or data associated with the cursor's current
601 * position. Any prior record or data stored in the cursor is replaced.
602 * The cursor must be positioned at a leaf node.
603 *
47197d71 604 * NOTE: All extractions occur at the leaf of the B-Tree.
8cd0a023 605 */
66325755 606int
8cd0a023 607hammer_btree_extract(hammer_cursor_t cursor, int flags)
66325755 608{
47197d71 609 hammer_mount_t hmp;
8cd0a023
MD
610 hammer_node_ondisk_t node;
611 hammer_btree_elm_t elm;
47197d71 612 hammer_off_t data_off;
19619882 613 int32_t data_len;
427e5fc6 614 int error;
427e5fc6 615
8cd0a023 616 /*
427e5fc6
MD
617 * The case where the data reference resolves to the same buffer
618 * as the record reference must be handled.
619 */
8cd0a023 620 node = cursor->node->ondisk;
8cd0a023 621 elm = &node->elms[cursor->index];
40043e7f
MD
622 cursor->data = NULL;
623 hmp = cursor->node->hmp;
66325755 624
d26d0ae9 625 /*
fe7678ee 626 * There is nothing to extract for an internal element.
d26d0ae9 627 */
fe7678ee
MD
628 if (node->type == HAMMER_BTREE_TYPE_INTERNAL)
629 return(EINVAL);
630
47197d71
MD
631 /*
632 * Only record types have data.
633 */
fe7678ee 634 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
11ad5ade 635 cursor->leaf = &elm->leaf;
4a2796f3
MD
636
637 if ((flags & HAMMER_CURSOR_GET_DATA) == 0)
638 return(0);
47197d71 639 if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
4a2796f3 640 return(0);
47197d71 641 data_off = elm->leaf.data_offset;
19619882 642 data_len = elm->leaf.data_len;
47197d71 643 if (data_off == 0)
4a2796f3 644 return(0);
d26d0ae9 645
4a2796f3
MD
646 /*
647 * Load the data
648 */
649 KKASSERT(data_len >= 0 && data_len <= HAMMER_XBUFSIZE);
650 cursor->data = hammer_bread_ext(hmp, data_off, data_len,
651 &error, &cursor->data_buffer);
ddfdf542 652 if (hammer_crc_test_leaf(cursor->data, &elm->leaf) == 0)
4a2796f3 653 Debugger("CRC FAILED: DATA");
427e5fc6
MD
654 return(error);
655}
656
657
658/*
8cd0a023
MD
659 * Insert a leaf element into the B-Tree at the current cursor position.
660 * The cursor is positioned such that the element at and beyond the cursor
661 * are shifted to make room for the new record.
662 *
a89aec1b 663 * The caller must call hammer_btree_lookup() with the HAMMER_CURSOR_INSERT
8cd0a023
MD
664 * flag set and that call must return ENOENT before this function can be
665 * called.
666 *
d36ec43b 667 * The caller may depend on the cursor's exclusive lock after return to
1f07f686 668 * interlock frontend visibility (see HAMMER_RECF_CONVERT_DELETE).
d36ec43b 669 *
8cd0a023 670 * ENOSPC is returned if there is no room to insert a new record.
427e5fc6
MD
671 */
672int
602c6cb8
MD
673hammer_btree_insert(hammer_cursor_t cursor, hammer_btree_leaf_elm_t elm,
674 int *doprop)
427e5fc6 675{
8cd0a023 676 hammer_node_ondisk_t node;
427e5fc6 677 int i;
6a37e7e4
MD
678 int error;
679
602c6cb8 680 *doprop = 0;
7bc5b8c2 681 if ((error = hammer_cursor_upgrade_node(cursor)) != 0)
6a37e7e4 682 return(error);
cb51be26 683 ++hammer_stats_btree_inserts;
427e5fc6 684
427e5fc6
MD
685 /*
686 * Insert the element at the leaf node and update the count in the
687 * parent. It is possible for parent to be NULL, indicating that
47197d71
MD
688 * the filesystem's ROOT B-Tree node is a leaf itself, which is
689 * possible. The root inode can never be deleted so the leaf should
690 * never be empty.
427e5fc6
MD
691 *
692 * Remember that the right-hand boundary is not included in the
693 * count.
694 */
36f82b23 695 hammer_modify_node_all(cursor->trans, cursor->node);
8cd0a023 696 node = cursor->node->ondisk;
427e5fc6 697 i = cursor->index;
fe7678ee 698 KKASSERT(elm->base.btype != 0);
8cd0a023
MD
699 KKASSERT(node->type == HAMMER_BTREE_TYPE_LEAF);
700 KKASSERT(node->count < HAMMER_BTREE_LEAF_ELMS);
701 if (i != node->count) {
702 bcopy(&node->elms[i], &node->elms[i+1],
703 (node->count - i) * sizeof(*elm));
704 }
11ad5ade 705 node->elms[i].leaf = *elm;
8cd0a023 706 ++node->count;
c82af904
MD
707
708 /*
709 * Update the leaf node's aggregate mirror_tid for mirroring
710 * support.
711 */
602c6cb8 712 if (node->mirror_tid < elm->base.delete_tid) {
c82af904 713 node->mirror_tid = elm->base.delete_tid;
602c6cb8
MD
714 *doprop = 1;
715 }
716 if (node->mirror_tid < elm->base.create_tid) {
c82af904 717 node->mirror_tid = elm->base.create_tid;
602c6cb8 718 *doprop = 1;
c82af904 719 }
602c6cb8 720 hammer_modify_node_done(cursor->node);
c82af904 721
eaeff70d 722 /*
47197d71 723 * Debugging sanity checks.
eaeff70d 724 */
11ad5ade
MD
725 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->base) <= 0);
726 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->base) > 0);
eaeff70d 727 if (i) {
11ad5ade 728 KKASSERT(hammer_btree_cmp(&node->elms[i-1].leaf.base, &elm->base) < 0);
eaeff70d 729 }
b3deaf57 730 if (i != node->count - 1)
11ad5ade 731 KKASSERT(hammer_btree_cmp(&node->elms[i+1].leaf.base, &elm->base) > 0);
b3deaf57 732
427e5fc6
MD
733 return(0);
734}
735
736/*
fe7678ee 737 * Delete a record from the B-Tree at the current cursor position.
8cd0a023
MD
738 * The cursor is positioned such that the current element is the one
739 * to be deleted.
740 *
195c19a1
MD
741 * On return the cursor will be positioned after the deleted element and
742 * MAY point to an internal node. It will be suitable for the continuation
743 * of an iteration but not for an insertion or deletion.
8cd0a023 744 *
195c19a1 745 * Deletions will attempt to partially rebalance the B-Tree in an upward
f36a9737
MD
746 * direction, but will terminate rather then deadlock. Empty internal nodes
747 * are never allowed by a deletion which deadlocks may end up giving us an
748 * empty leaf. The pruner will clean up and rebalance the tree.
46fe7ae1
MD
749 *
750 * This function can return EDEADLK, requiring the caller to retry the
751 * operation after clearing the deadlock.
427e5fc6
MD
752 */
753int
8cd0a023 754hammer_btree_delete(hammer_cursor_t cursor)
427e5fc6 755{
8cd0a023
MD
756 hammer_node_ondisk_t ondisk;
757 hammer_node_t node;
758 hammer_node_t parent;
8cd0a023 759 int error;
427e5fc6
MD
760 int i;
761
6a37e7e4
MD
762 if ((error = hammer_cursor_upgrade(cursor)) != 0)
763 return(error);
cb51be26 764 ++hammer_stats_btree_deletes;
6a37e7e4 765
427e5fc6 766 /*
8cd0a023 767 * Delete the element from the leaf node.
427e5fc6 768 *
8cd0a023 769 * Remember that leaf nodes do not have boundaries.
427e5fc6 770 */
8cd0a023
MD
771 node = cursor->node;
772 ondisk = node->ondisk;
427e5fc6
MD
773 i = cursor->index;
774
8cd0a023 775 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_LEAF);
fe7678ee 776 KKASSERT(i >= 0 && i < ondisk->count);
36f82b23 777 hammer_modify_node_all(cursor->trans, node);
8cd0a023
MD
778 if (i + 1 != ondisk->count) {
779 bcopy(&ondisk->elms[i+1], &ondisk->elms[i],
780 (ondisk->count - i - 1) * sizeof(ondisk->elms[0]));
781 }
782 --ondisk->count;
10a5d1ba 783 hammer_modify_node_done(node);
b3bad96f 784 hammer_cursor_deleted_element(node, i);
fe7678ee
MD
785
786 /*
787 * Validate local parent
788 */
789 if (ondisk->parent) {
8cd0a023 790 parent = cursor->parent;
fe7678ee
MD
791
792 KKASSERT(parent != NULL);
793 KKASSERT(parent->node_offset == ondisk->parent);
427e5fc6 794 }
427e5fc6 795
8cd0a023 796 /*
fe7678ee 797 * If the leaf becomes empty it must be detached from the parent,
47197d71 798 * potentially recursing through to the filesystem root.
195c19a1
MD
799 *
800 * This may reposition the cursor at one of the parent's of the
801 * current node.
6a37e7e4
MD
802 *
803 * Ignore deadlock errors, that simply means that btree_remove
f36a9737 804 * was unable to recurse and had to leave us with an empty leaf.
8cd0a023 805 */
b3deaf57 806 KKASSERT(cursor->index <= ondisk->count);
8cd0a023 807 if (ondisk->count == 0) {
f36a9737 808 error = btree_remove(cursor);
6a37e7e4
MD
809 if (error == EDEADLK)
810 error = 0;
8cd0a023 811 } else {
8cd0a023
MD
812 error = 0;
813 }
eaeff70d
MD
814 KKASSERT(cursor->parent == NULL ||
815 cursor->parent_index < cursor->parent->ondisk->count);
8cd0a023
MD
816 return(error);
817}
427e5fc6
MD
818
819/*
8cd0a023
MD
820 * PRIMAY B-TREE SEARCH SUPPORT PROCEDURE
821 *
47197d71 822 * Search the filesystem B-Tree for cursor->key_beg, return the matching node.
8cd0a023 823 *
d26d0ae9
MD
824 * The search can begin ANYWHERE in the B-Tree. As a first step the search
825 * iterates up the tree as necessary to properly position itself prior to
826 * actually doing the sarch.
827 *
8cd0a023 828 * INSERTIONS: The search will split full nodes and leaves on its way down
d26d0ae9
MD
829 * and guarentee that the leaf it ends up on is not full. If we run out
830 * of space the search continues to the leaf (to position the cursor for
831 * the spike), but ENOSPC is returned.
427e5fc6 832 *
fbc6e32a
MD
833 * The search is only guarenteed to end up on a leaf if an error code of 0
834 * is returned, or if inserting and an error code of ENOENT is returned.
d26d0ae9 835 * Otherwise it can stop at an internal node. On success a search returns
47197d71 836 * a leaf node.
eaeff70d
MD
837 *
838 * COMPLEXITY WARNING! This is the core B-Tree search code for the entire
839 * filesystem, and it is not simple code. Please note the following facts:
840 *
841 * - Internal node recursions have a boundary on the left AND right. The
9582c7da 842 * right boundary is non-inclusive. The create_tid is a generic part
eaeff70d
MD
843 * of the key for internal nodes.
844 *
47197d71 845 * - Leaf nodes contain terminal elements only now.
eaeff70d
MD
846 *
847 * - Filesystem lookups typically set HAMMER_CURSOR_ASOF, indicating a
b33e2cc0
MD
848 * historical search. ASOF and INSERT are mutually exclusive. When
849 * doing an as-of lookup btree_search() checks for a right-edge boundary
9582c7da
MD
850 * case. If while recursing down the left-edge differs from the key
851 * by ONLY its create_tid, HAMMER_CURSOR_CREATE_CHECK is set along
852 * with cursor->create_check. This is used by btree_lookup() to iterate.
853 * The iteration backwards because as-of searches can wind up going
b33e2cc0 854 * down the wrong branch of the B-Tree.
427e5fc6 855 */
8cd0a023 856static
427e5fc6 857int
8cd0a023 858btree_search(hammer_cursor_t cursor, int flags)
427e5fc6 859{
8cd0a023 860 hammer_node_ondisk_t node;
61aeeb33 861 hammer_btree_elm_t elm;
8cd0a023 862 int error;
d26d0ae9 863 int enospc = 0;
8cd0a023
MD
864 int i;
865 int r;
b33e2cc0 866 int s;
8cd0a023
MD
867
868 flags |= cursor->flags;
cb51be26 869 ++hammer_stats_btree_searches;
8cd0a023 870
b3deaf57 871 if (hammer_debug_btree) {
2f85fa4d 872 kprintf("SEARCH %016llx[%d] %016llx %02x key=%016llx cre=%016llx lo=%02x (td = %p)\n",
eaeff70d
MD
873 cursor->node->node_offset,
874 cursor->index,
b3deaf57
MD
875 cursor->key_beg.obj_id,
876 cursor->key_beg.rec_type,
d113fda1 877 cursor->key_beg.key,
a84a197d 878 cursor->key_beg.create_tid,
2f85fa4d 879 cursor->key_beg.localization,
a84a197d 880 curthread
b3deaf57 881 );
a84a197d
MD
882 if (cursor->parent)
883 kprintf("SEARCHP %016llx[%d] (%016llx/%016llx %016llx/%016llx) (%p/%p %p/%p)\n",
884 cursor->parent->node_offset, cursor->parent_index,
885 cursor->left_bound->obj_id,
886 cursor->parent->ondisk->elms[cursor->parent_index].internal.base.obj_id,
887 cursor->right_bound->obj_id,
888 cursor->parent->ondisk->elms[cursor->parent_index+1].internal.base.obj_id,
889 cursor->left_bound,
890 &cursor->parent->ondisk->elms[cursor->parent_index],
891 cursor->right_bound,
892 &cursor->parent->ondisk->elms[cursor->parent_index+1]
893 );
b3deaf57
MD
894 }
895
8cd0a023
MD
896 /*
897 * Move our cursor up the tree until we find a node whos range covers
47197d71 898 * the key we are trying to locate.
8cd0a023
MD
899 *
900 * The left bound is inclusive, the right bound is non-inclusive.
47197d71 901 * It is ok to cursor up too far.
8cd0a023 902 */
b33e2cc0
MD
903 for (;;) {
904 r = hammer_btree_cmp(&cursor->key_beg, cursor->left_bound);
905 s = hammer_btree_cmp(&cursor->key_beg, cursor->right_bound);
906 if (r >= 0 && s < 0)
907 break;
9944ae54 908 KKASSERT(cursor->parent);
cb51be26 909 ++hammer_stats_btree_iterations;
6a37e7e4 910 error = hammer_cursor_up(cursor);
8cd0a023
MD
911 if (error)
912 goto done;
427e5fc6 913 }
427e5fc6 914
b33e2cc0
MD
915 /*
916 * The delete-checks below are based on node, not parent. Set the
917 * initial delete-check based on the parent.
918 */
9582c7da
MD
919 if (r == 1) {
920 KKASSERT(cursor->left_bound->create_tid != 1);
921 cursor->create_check = cursor->left_bound->create_tid - 1;
922 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
b33e2cc0
MD
923 }
924
8cd0a023 925 /*
47197d71 926 * We better have ended up with a node somewhere.
8cd0a023 927 */
47197d71 928 KKASSERT(cursor->node != NULL);
8cd0a023
MD
929
930 /*
931 * If we are inserting we can't start at a full node if the parent
932 * is also full (because there is no way to split the node),
b33e2cc0 933 * continue running up the tree until the requirement is satisfied
47197d71 934 * or we hit the root of the filesystem.
9582c7da
MD
935 *
936 * (If inserting we aren't doing an as-of search so we don't have
937 * to worry about create_check).
8cd0a023 938 */
61aeeb33 939 while ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
eaeff70d
MD
940 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
941 if (btree_node_is_full(cursor->node->ondisk) == 0)
942 break;
943 } else {
47197d71 944 if (btree_node_is_full(cursor->node->ondisk) ==0)
eaeff70d
MD
945 break;
946 }
b33e2cc0
MD
947 if (cursor->node->ondisk->parent == 0 ||
948 cursor->parent->ondisk->count != HAMMER_BTREE_INT_ELMS) {
8cd0a023 949 break;
b33e2cc0 950 }
cb51be26 951 ++hammer_stats_btree_iterations;
6a37e7e4 952 error = hammer_cursor_up(cursor);
47197d71 953 /* node may have become stale */
8cd0a023
MD
954 if (error)
955 goto done;
427e5fc6 956 }
427e5fc6 957
8cd0a023
MD
958 /*
959 * Push down through internal nodes to locate the requested key.
960 */
8cd0a023
MD
961 node = cursor->node->ondisk;
962 while (node->type == HAMMER_BTREE_TYPE_INTERNAL) {
8cd0a023
MD
963 /*
964 * Scan the node to find the subtree index to push down into.
fbc6e32a 965 * We go one-past, then back-up.
d113fda1 966 *
fe7678ee
MD
967 * We must proactively remove deleted elements which may
968 * have been left over from a deadlocked btree_remove().
969 *
eaeff70d 970 * The left and right boundaries are included in the loop
d5530d22 971 * in order to detect edge cases.
9944ae54 972 *
9582c7da 973 * If the separator only differs by create_tid (r == 1)
eaeff70d
MD
974 * and we are doing an as-of search, we may end up going
975 * down a branch to the left of the one containing the
976 * desired key. This requires numerous special cases.
8cd0a023 977 */
47637bff 978 ++hammer_stats_btree_iterations;
46fe7ae1 979 if (hammer_debug_btree) {
47197d71 980 kprintf("SEARCH-I %016llx count=%d\n",
46fe7ae1
MD
981 cursor->node->node_offset,
982 node->count);
983 }
af209b0f
MD
984
985 /*
986 * Try to shortcut the search before dropping into the
987 * linear loop. Locate the first node where r <= 1.
988 */
989 i = hammer_btree_search_node(&cursor->key_beg, node);
990 while (i <= node->count) {
cb51be26 991 ++hammer_stats_btree_elements;
61aeeb33
MD
992 elm = &node->elms[i];
993 r = hammer_btree_cmp(&cursor->key_beg, &elm->base);
b33e2cc0
MD
994 if (hammer_debug_btree > 2) {
995 kprintf(" IELM %p %d r=%d\n",
996 &node->elms[i], i, r);
997 }
9582c7da 998 if (r < 0)
8cd0a023 999 break;
9582c7da
MD
1000 if (r == 1) {
1001 KKASSERT(elm->base.create_tid != 1);
1002 cursor->create_check = elm->base.create_tid - 1;
1003 cursor->flags |= HAMMER_CURSOR_CREATE_CHECK;
b33e2cc0 1004 }
af209b0f 1005 ++i;
8cd0a023 1006 }
eaeff70d 1007 if (hammer_debug_btree) {
46fe7ae1
MD
1008 kprintf("SEARCH-I preI=%d/%d r=%d\n",
1009 i, node->count, r);
eaeff70d 1010 }
8cd0a023
MD
1011
1012 /*
9944ae54
MD
1013 * These cases occur when the parent's idea of the boundary
1014 * is wider then the child's idea of the boundary, and
1015 * require special handling. If not inserting we can
1016 * terminate the search early for these cases but the
1017 * child's boundaries cannot be unconditionally modified.
8cd0a023 1018 */
fbc6e32a 1019 if (i == 0) {
9944ae54
MD
1020 /*
1021 * If i == 0 the search terminated to the LEFT of the
1022 * left_boundary but to the RIGHT of the parent's left
1023 * boundary.
1024 */
fbc6e32a 1025 u_int8_t save;
d26d0ae9 1026
eaeff70d
MD
1027 elm = &node->elms[0];
1028
1029 /*
1030 * If we aren't inserting we can stop here.
1031 */
11ad5ade
MD
1032 if ((flags & (HAMMER_CURSOR_INSERT |
1033 HAMMER_CURSOR_PRUNING)) == 0) {
fbc6e32a
MD
1034 cursor->index = 0;
1035 return(ENOENT);
1036 }
9944ae54 1037
d5530d22
MD
1038 /*
1039 * Correct a left-hand boundary mismatch.
6a37e7e4 1040 *
f36a9737
MD
1041 * We can only do this if we can upgrade the lock,
1042 * and synchronized as a background cursor (i.e.
1043 * inserting or pruning).
10a5d1ba
MD
1044 *
1045 * WARNING: We can only do this if inserting, i.e.
1046 * we are running on the backend.
d5530d22 1047 */
eaeff70d
MD
1048 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1049 return(error);
10a5d1ba 1050 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
c9b9e29d
MD
1051 hammer_modify_node_field(cursor->trans, cursor->node,
1052 elms[0]);
fe7678ee 1053 save = node->elms[0].base.btype;
d5530d22 1054 node->elms[0].base = *cursor->left_bound;
fe7678ee 1055 node->elms[0].base.btype = save;
10a5d1ba 1056 hammer_modify_node_done(cursor->node);
9944ae54 1057 } else if (i == node->count + 1) {
d26d0ae9 1058 /*
9944ae54
MD
1059 * If i == node->count + 1 the search terminated to
1060 * the RIGHT of the right boundary but to the LEFT
eaeff70d
MD
1061 * of the parent's right boundary. If we aren't
1062 * inserting we can stop here.
d113fda1 1063 *
9944ae54
MD
1064 * Note that the last element in this case is
1065 * elms[i-2] prior to adjustments to 'i'.
d26d0ae9 1066 */
9944ae54 1067 --i;
11ad5ade
MD
1068 if ((flags & (HAMMER_CURSOR_INSERT |
1069 HAMMER_CURSOR_PRUNING)) == 0) {
9944ae54 1070 cursor->index = i;
eaeff70d 1071 return (ENOENT);
d26d0ae9
MD
1072 }
1073
d5530d22
MD
1074 /*
1075 * Correct a right-hand boundary mismatch.
1076 * (actual push-down record is i-2 prior to
1077 * adjustments to i).
6a37e7e4 1078 *
f36a9737
MD
1079 * We can only do this if we can upgrade the lock,
1080 * and synchronized as a background cursor (i.e.
1081 * inserting or pruning).
10a5d1ba
MD
1082 *
1083 * WARNING: We can only do this if inserting, i.e.
1084 * we are running on the backend.
d5530d22 1085 */
eaeff70d
MD
1086 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1087 return(error);
9944ae54 1088 elm = &node->elms[i];
10a5d1ba 1089 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
36f82b23
MD
1090 hammer_modify_node(cursor->trans, cursor->node,
1091 &elm->base, sizeof(elm->base));
d5530d22 1092 elm->base = *cursor->right_bound;
10a5d1ba 1093 hammer_modify_node_done(cursor->node);
d5530d22 1094 --i;
fbc6e32a
MD
1095 } else {
1096 /*
9944ae54
MD
1097 * The push-down index is now i - 1. If we had
1098 * terminated on the right boundary this will point
1099 * us at the last element.
fbc6e32a
MD
1100 */
1101 --i;
1102 }
8cd0a023 1103 cursor->index = i;
6a37e7e4 1104 elm = &node->elms[i];
8cd0a023 1105
b3deaf57 1106 if (hammer_debug_btree) {
47197d71 1107 kprintf("RESULT-I %016llx[%d] %016llx %02x "
2f85fa4d 1108 "key=%016llx cre=%016llx lo=%02x\n",
eaeff70d
MD
1109 cursor->node->node_offset,
1110 i,
b3deaf57
MD
1111 elm->internal.base.obj_id,
1112 elm->internal.base.rec_type,
d113fda1 1113 elm->internal.base.key,
2f85fa4d
MD
1114 elm->internal.base.create_tid,
1115 elm->internal.base.localization
b3deaf57
MD
1116 );
1117 }
1118
6a37e7e4 1119 /*
f36a9737 1120 * We better have a valid subtree offset.
6a37e7e4 1121 */
f36a9737 1122 KKASSERT(elm->internal.subtree_offset != 0);
6a37e7e4 1123
8cd0a023
MD
1124 /*
1125 * Handle insertion and deletion requirements.
1126 *
1127 * If inserting split full nodes. The split code will
1128 * adjust cursor->node and cursor->index if the current
1129 * index winds up in the new node.
61aeeb33 1130 *
9944ae54
MD
1131 * If inserting and a left or right edge case was detected,
1132 * we cannot correct the left or right boundary and must
1133 * prepend and append an empty leaf node in order to make
1134 * the boundary correction.
1135 *
61aeeb33
MD
1136 * If we run out of space we set enospc and continue on
1137 * to a leaf to provide the spike code with a good point
47197d71 1138 * of entry.
8cd0a023 1139 */
61aeeb33 1140 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0) {
fe7678ee 1141 if (btree_node_is_full(node)) {
8cd0a023 1142 error = btree_split_internal(cursor);
d26d0ae9
MD
1143 if (error) {
1144 if (error != ENOSPC)
1145 goto done;
1146 enospc = 1;
d26d0ae9 1147 }
8cd0a023
MD
1148 /*
1149 * reload stale pointers
1150 */
1151 i = cursor->index;
1152 node = cursor->node->ondisk;
1153 }
d26d0ae9 1154 }
427e5fc6
MD
1155
1156 /*
8cd0a023 1157 * Push down (push into new node, existing node becomes
d26d0ae9 1158 * the parent) and continue the search.
427e5fc6 1159 */
8cd0a023 1160 error = hammer_cursor_down(cursor);
47197d71 1161 /* node may have become stale */
8cd0a023
MD
1162 if (error)
1163 goto done;
1164 node = cursor->node->ondisk;
427e5fc6 1165 }
427e5fc6 1166
8cd0a023
MD
1167 /*
1168 * We are at a leaf, do a linear search of the key array.
d26d0ae9
MD
1169 *
1170 * On success the index is set to the matching element and 0
1171 * is returned.
1172 *
1173 * On failure the index is set to the insertion point and ENOENT
1174 * is returned.
8cd0a023
MD
1175 *
1176 * Boundaries are not stored in leaf nodes, so the index can wind
1177 * up to the left of element 0 (index == 0) or past the end of
f36a9737
MD
1178 * the array (index == node->count). It is also possible that the
1179 * leaf might be empty.
8cd0a023 1180 */
47637bff 1181 ++hammer_stats_btree_iterations;
fe7678ee 1182 KKASSERT (node->type == HAMMER_BTREE_TYPE_LEAF);
8cd0a023 1183 KKASSERT(node->count <= HAMMER_BTREE_LEAF_ELMS);
46fe7ae1 1184 if (hammer_debug_btree) {
47197d71 1185 kprintf("SEARCH-L %016llx count=%d\n",
46fe7ae1
MD
1186 cursor->node->node_offset,
1187 node->count);
1188 }
8cd0a023 1189
af209b0f
MD
1190 /*
1191 * Try to shortcut the search before dropping into the
1192 * linear loop. Locate the first node where r <= 1.
1193 */
1194 i = hammer_btree_search_node(&cursor->key_beg, node);
1195 while (i < node->count) {
cb51be26 1196 ++hammer_stats_btree_elements;
fe7678ee
MD
1197 elm = &node->elms[i];
1198
1199 r = hammer_btree_cmp(&cursor->key_beg, &elm->leaf.base);
427e5fc6 1200
d5530d22
MD
1201 if (hammer_debug_btree > 1)
1202 kprintf(" ELM %p %d r=%d\n", &node->elms[i], i, r);
1203
427e5fc6 1204 /*
fe7678ee 1205 * We are at a record element. Stop if we've flipped past
9582c7da
MD
1206 * key_beg, not counting the create_tid test. Allow the
1207 * r == 1 case (key_beg > element but differs only by its
1208 * create_tid) to fall through to the AS-OF check.
427e5fc6 1209 */
fe7678ee
MD
1210 KKASSERT (elm->leaf.base.btype == HAMMER_BTREE_TYPE_RECORD);
1211
9582c7da 1212 if (r < 0)
d5530d22 1213 goto failed;
af209b0f
MD
1214 if (r > 1) {
1215 ++i;
d5530d22 1216 continue;
af209b0f 1217 }
427e5fc6 1218
66325755 1219 /*
9582c7da 1220 * Check our as-of timestamp against the element.
66325755 1221 */
eaeff70d 1222 if (flags & HAMMER_CURSOR_ASOF) {
fe7678ee 1223 if (hammer_btree_chkts(cursor->asof,
d113fda1 1224 &node->elms[i].base) != 0) {
af209b0f 1225 ++i;
d113fda1
MD
1226 continue;
1227 }
eaeff70d
MD
1228 /* success */
1229 } else {
af209b0f
MD
1230 if (r > 0) { /* can only be +1 */
1231 ++i;
9582c7da 1232 continue;
af209b0f 1233 }
eaeff70d 1234 /* success */
66325755 1235 }
d5530d22
MD
1236 cursor->index = i;
1237 error = 0;
eaeff70d 1238 if (hammer_debug_btree) {
47197d71
MD
1239 kprintf("RESULT-L %016llx[%d] (SUCCESS)\n",
1240 cursor->node->node_offset, i);
eaeff70d 1241 }
d5530d22
MD
1242 goto done;
1243 }
1244
1245 /*
eaeff70d 1246 * The search of the leaf node failed. i is the insertion point.
d5530d22 1247 */
d5530d22 1248failed:
b3deaf57 1249 if (hammer_debug_btree) {
47197d71
MD
1250 kprintf("RESULT-L %016llx[%d] (FAILED)\n",
1251 cursor->node->node_offset, i);
b3deaf57
MD
1252 }
1253
8cd0a023
MD
1254 /*
1255 * No exact match was found, i is now at the insertion point.
1256 *
1257 * If inserting split a full leaf before returning. This
1258 * may have the side effect of adjusting cursor->node and
1259 * cursor->index.
1260 */
1261 cursor->index = i;
eaeff70d 1262 if ((flags & HAMMER_CURSOR_INSERT) && enospc == 0 &&
47197d71 1263 btree_node_is_full(node)) {
8cd0a023 1264 error = btree_split_leaf(cursor);
d26d0ae9
MD
1265 if (error) {
1266 if (error != ENOSPC)
1267 goto done;
1268 enospc = 1;
d26d0ae9
MD
1269 }
1270 /*
1271 * reload stale pointers
1272 */
8cd0a023
MD
1273 /* NOT USED
1274 i = cursor->index;
1275 node = &cursor->node->internal;
1276 */
8cd0a023 1277 }
d26d0ae9
MD
1278
1279 /*
1280 * We reached a leaf but did not find the key we were looking for.
1281 * If this is an insert we will be properly positioned for an insert
1282 * (ENOENT) or spike (ENOSPC) operation.
1283 */
1284 error = enospc ? ENOSPC : ENOENT;
8cd0a023 1285done:
427e5fc6
MD
1286 return(error);
1287}
1288
af209b0f
MD
1289/*
1290 * Heuristical search for the first element whos comparison is <= 1. May
1291 * return an index whos compare result is > 1 but may only return an index
1292 * whos compare result is <= 1 if it is the first element with that result.
1293 */
bcac4bbb 1294int
af209b0f
MD
1295hammer_btree_search_node(hammer_base_elm_t elm, hammer_node_ondisk_t node)
1296{
1297 int b;
1298 int s;
1299 int i;
1300 int r;
1301
1302 /*
1303 * Don't bother if the node does not have very many elements
1304 */
1305 b = 0;
1306 s = node->count;
1307 while (s - b > 4) {
1308 i = b + (s - b) / 2;
cb51be26 1309 ++hammer_stats_btree_elements;
af209b0f
MD
1310 r = hammer_btree_cmp(elm, &node->elms[i].leaf.base);
1311 if (r <= 1) {
1312 s = i;
1313 } else {
1314 b = i;
1315 }
1316 }
1317 return(b);
1318}
1319
8cd0a023 1320
427e5fc6 1321/************************************************************************
8cd0a023 1322 * SPLITTING AND MERGING *
427e5fc6
MD
1323 ************************************************************************
1324 *
1325 * These routines do all the dirty work required to split and merge nodes.
1326 */
1327
1328/*
8cd0a023 1329 * Split an internal node into two nodes and move the separator at the split
fe7678ee 1330 * point to the parent.
427e5fc6 1331 *
8cd0a023
MD
1332 * (cursor->node, cursor->index) indicates the element the caller intends
1333 * to push into. We will adjust node and index if that element winds
427e5fc6 1334 * up in the split node.
8cd0a023 1335 *
47197d71
MD
1336 * If we are at the root of the filesystem a new root must be created with
1337 * two elements, one pointing to the original root and one pointing to the
8cd0a023 1338 * newly allocated split node.
427e5fc6
MD
1339 */
1340static
1341int
8cd0a023 1342btree_split_internal(hammer_cursor_t cursor)
427e5fc6 1343{
8cd0a023
MD
1344 hammer_node_ondisk_t ondisk;
1345 hammer_node_t node;
1346 hammer_node_t parent;
1347 hammer_node_t new_node;
1348 hammer_btree_elm_t elm;
1349 hammer_btree_elm_t parent_elm;
b33e2cc0 1350 hammer_node_locklist_t locklist = NULL;
36f82b23 1351 hammer_mount_t hmp = cursor->trans->hmp;
427e5fc6
MD
1352 int parent_index;
1353 int made_root;
1354 int split;
1355 int error;
7f7c1f84 1356 int i;
8cd0a023 1357 const int esize = sizeof(*elm);
427e5fc6 1358
47197d71
MD
1359 error = hammer_btree_lock_children(cursor, &locklist);
1360 if (error)
1361 goto done;
7bc5b8c2
MD
1362 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1363 goto done;
cb51be26 1364 ++hammer_stats_btree_splits;
6a37e7e4 1365
427e5fc6
MD
1366 /*
1367 * We are splitting but elms[split] will be promoted to the parent,
1368 * leaving the right hand node with one less element. If the
1369 * insertion point will be on the left-hand side adjust the split
1370 * point to give the right hand side one additional node.
1371 */
8cd0a023
MD
1372 node = cursor->node;
1373 ondisk = node->ondisk;
1374 split = (ondisk->count + 1) / 2;
427e5fc6
MD
1375 if (cursor->index <= split)
1376 --split;
427e5fc6
MD
1377
1378 /*
47197d71
MD
1379 * If we are at the root of the filesystem, create a new root node
1380 * with 1 element and split normally. Avoid making major
1381 * modifications until we know the whole operation will work.
427e5fc6 1382 */
8cd0a023 1383 if (ondisk->parent == 0) {
36f82b23 1384 parent = hammer_alloc_btree(cursor->trans, &error);
427e5fc6 1385 if (parent == NULL)
6a37e7e4 1386 goto done;
8cd0a023 1387 hammer_lock_ex(&parent->lock);
36f82b23 1388 hammer_modify_node_noundo(cursor->trans, parent);
8cd0a023
MD
1389 ondisk = parent->ondisk;
1390 ondisk->count = 1;
1391 ondisk->parent = 0;
1392 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
47197d71 1393 ondisk->elms[0].base = hmp->root_btree_beg;
fe7678ee 1394 ondisk->elms[0].base.btype = node->ondisk->type;
8cd0a023 1395 ondisk->elms[0].internal.subtree_offset = node->node_offset;
47197d71 1396 ondisk->elms[1].base = hmp->root_btree_end;
10a5d1ba 1397 hammer_modify_node_done(parent);
fe7678ee 1398 /* ondisk->elms[1].base.btype - not used */
427e5fc6 1399 made_root = 1;
8cd0a023 1400 parent_index = 0; /* index of current node in parent */
427e5fc6
MD
1401 } else {
1402 made_root = 0;
8cd0a023
MD
1403 parent = cursor->parent;
1404 parent_index = cursor->parent_index;
427e5fc6 1405 }
427e5fc6
MD
1406
1407 /*
1408 * Split node into new_node at the split point.
1409 *
1410 * B O O O P N N B <-- P = node->elms[split]
1411 * 0 1 2 3 4 5 6 <-- subtree indices
1412 *
1413 * x x P x x
1414 * s S S s
1415 * / \
1416 * B O O O B B N N B <--- inner boundary points are 'P'
1417 * 0 1 2 3 4 5 6
1418 *
1419 */
36f82b23 1420 new_node = hammer_alloc_btree(cursor->trans, &error);
427e5fc6 1421 if (new_node == NULL) {
8cd0a023
MD
1422 if (made_root) {
1423 hammer_unlock(&parent->lock);
36f82b23 1424 hammer_delete_node(cursor->trans, parent);
8cd0a023
MD
1425 hammer_rel_node(parent);
1426 }
6a37e7e4 1427 goto done;
427e5fc6 1428 }
8cd0a023 1429 hammer_lock_ex(&new_node->lock);
427e5fc6
MD
1430
1431 /*
8cd0a023 1432 * Create the new node. P becomes the left-hand boundary in the
427e5fc6
MD
1433 * new node. Copy the right-hand boundary as well.
1434 *
1435 * elm is the new separator.
1436 */
36f82b23
MD
1437 hammer_modify_node_noundo(cursor->trans, new_node);
1438 hammer_modify_node_all(cursor->trans, node);
8cd0a023
MD
1439 ondisk = node->ondisk;
1440 elm = &ondisk->elms[split];
1441 bcopy(elm, &new_node->ondisk->elms[0],
1442 (ondisk->count - split + 1) * esize);
1443 new_node->ondisk->count = ondisk->count - split;
1444 new_node->ondisk->parent = parent->node_offset;
1445 new_node->ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
1446 KKASSERT(ondisk->type == new_node->ondisk->type);
b3bad96f 1447 hammer_cursor_split_node(node, new_node, split);
427e5fc6
MD
1448
1449 /*
fe7678ee
MD
1450 * Cleanup the original node. Elm (P) becomes the new boundary,
1451 * its subtree_offset was moved to the new node. If we had created
427e5fc6
MD
1452 * a new root its parent pointer may have changed.
1453 */
8cd0a023 1454 elm->internal.subtree_offset = 0;
c0ade690 1455 ondisk->count = split;
427e5fc6
MD
1456
1457 /*
1458 * Insert the separator into the parent, fixup the parent's
1459 * reference to the original node, and reference the new node.
1460 * The separator is P.
1461 *
1462 * Remember that base.count does not include the right-hand boundary.
1463 */
36f82b23 1464 hammer_modify_node_all(cursor->trans, parent);
8cd0a023 1465 ondisk = parent->ondisk;
d26d0ae9 1466 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
8cd0a023 1467 parent_elm = &ondisk->elms[parent_index+1];
427e5fc6 1468 bcopy(parent_elm, parent_elm + 1,
8cd0a023
MD
1469 (ondisk->count - parent_index) * esize);
1470 parent_elm->internal.base = elm->base; /* separator P */
fe7678ee 1471 parent_elm->internal.base.btype = new_node->ondisk->type;
8cd0a023 1472 parent_elm->internal.subtree_offset = new_node->node_offset;
76376933 1473 ++ondisk->count;
10a5d1ba 1474 hammer_modify_node_done(parent);
b3bad96f 1475 hammer_cursor_inserted_element(parent, parent_index + 1);
427e5fc6 1476
7f7c1f84
MD
1477 /*
1478 * The children of new_node need their parent pointer set to new_node.
b33e2cc0
MD
1479 * The children have already been locked by
1480 * hammer_btree_lock_children().
7f7c1f84
MD
1481 */
1482 for (i = 0; i < new_node->ondisk->count; ++i) {
1483 elm = &new_node->ondisk->elms[i];
36f82b23 1484 error = btree_set_parent(cursor->trans, new_node, elm);
7f7c1f84
MD
1485 if (error) {
1486 panic("btree_split_internal: btree-fixup problem");
1487 }
1488 }
10a5d1ba 1489 hammer_modify_node_done(new_node);
7f7c1f84 1490
427e5fc6 1491 /*
47197d71 1492 * The filesystem's root B-Tree pointer may have to be updated.
427e5fc6
MD
1493 */
1494 if (made_root) {
47197d71
MD
1495 hammer_volume_t volume;
1496
1497 volume = hammer_get_root_volume(hmp, &error);
1498 KKASSERT(error == 0);
1499
e8599db1
MD
1500 hammer_modify_volume_field(cursor->trans, volume,
1501 vol0_btree_root);
47197d71 1502 volume->ondisk->vol0_btree_root = parent->node_offset;
10a5d1ba 1503 hammer_modify_volume_done(volume);
8cd0a023
MD
1504 node->ondisk->parent = parent->node_offset;
1505 if (cursor->parent) {
1506 hammer_unlock(&cursor->parent->lock);
1507 hammer_rel_node(cursor->parent);
1508 }
1509 cursor->parent = parent; /* lock'd and ref'd */
47197d71 1510 hammer_rel_volume(volume, 0);
427e5fc6 1511 }
10a5d1ba 1512 hammer_modify_node_done(node);
427e5fc6
MD
1513
1514 /*
1515 * Ok, now adjust the cursor depending on which element the original
1516 * index was pointing at. If we are >= the split point the push node
1517 * is now in the new node.
1518 *
1519 * NOTE: If we are at the split point itself we cannot stay with the
1520 * original node because the push index will point at the right-hand
1521 * boundary, which is illegal.
8cd0a023
MD
1522 *
1523 * NOTE: The cursor's parent or parent_index must be adjusted for
1524 * the case where a new parent (new root) was created, and the case
1525 * where the cursor is now pointing at the split node.
427e5fc6
MD
1526 */
1527 if (cursor->index >= split) {
8cd0a023 1528 cursor->parent_index = parent_index + 1;
427e5fc6 1529 cursor->index -= split;
8cd0a023
MD
1530 hammer_unlock(&cursor->node->lock);
1531 hammer_rel_node(cursor->node);
1532 cursor->node = new_node; /* locked and ref'd */
1533 } else {
1534 cursor->parent_index = parent_index;
1535 hammer_unlock(&new_node->lock);
1536 hammer_rel_node(new_node);
427e5fc6 1537 }
76376933
MD
1538
1539 /*
1540 * Fixup left and right bounds
1541 */
1542 parent_elm = &parent->ondisk->elms[cursor->parent_index];
fbc6e32a
MD
1543 cursor->left_bound = &parent_elm[0].internal.base;
1544 cursor->right_bound = &parent_elm[1].internal.base;
b3deaf57
MD
1545 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1546 &cursor->node->ondisk->elms[0].internal.base) <= 0);
1547 KKASSERT(hammer_btree_cmp(cursor->right_bound,
9944ae54 1548 &cursor->node->ondisk->elms[cursor->node->ondisk->count].internal.base) >= 0);
76376933 1549
6a37e7e4 1550done:
b33e2cc0 1551 hammer_btree_unlock_children(&locklist);
6a37e7e4
MD
1552 hammer_cursor_downgrade(cursor);
1553 return (error);
427e5fc6
MD
1554}
1555
1556/*
1557 * Same as the above, but splits a full leaf node.
6a37e7e4
MD
1558 *
1559 * This function
427e5fc6
MD
1560 */
1561static
1562int
8cd0a023 1563btree_split_leaf(hammer_cursor_t cursor)
427e5fc6 1564{
8cd0a023
MD
1565 hammer_node_ondisk_t ondisk;
1566 hammer_node_t parent;
1567 hammer_node_t leaf;
47197d71 1568 hammer_mount_t hmp;
8cd0a023
MD
1569 hammer_node_t new_leaf;
1570 hammer_btree_elm_t elm;
1571 hammer_btree_elm_t parent_elm;
b3deaf57 1572 hammer_base_elm_t mid_boundary;
427e5fc6
MD
1573 int parent_index;
1574 int made_root;
1575 int split;
1576 int error;
8cd0a023 1577 const size_t esize = sizeof(*elm);
427e5fc6 1578
6a37e7e4
MD
1579 if ((error = hammer_cursor_upgrade(cursor)) != 0)
1580 return(error);
cb51be26 1581 ++hammer_stats_btree_splits;
6a37e7e4 1582
36f82b23
MD
1583 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1584 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1585 KKASSERT(hammer_btree_cmp(cursor->right_bound,
1586 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
1587
427e5fc6 1588 /*
8cd0a023
MD
1589 * Calculate the split point. If the insertion point will be on
1590 * the left-hand side adjust the split point to give the right
1591 * hand side one additional node.
fe7678ee
MD
1592 *
1593 * Spikes are made up of two leaf elements which cannot be
1594 * safely split.
427e5fc6 1595 */
8cd0a023
MD
1596 leaf = cursor->node;
1597 ondisk = leaf->ondisk;
1598 split = (ondisk->count + 1) / 2;
427e5fc6
MD
1599 if (cursor->index <= split)
1600 --split;
1601 error = 0;
40043e7f 1602 hmp = leaf->hmp;
427e5fc6 1603
fe7678ee 1604 elm = &ondisk->elms[split];
fe7678ee 1605
36f82b23
MD
1606 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm[-1].leaf.base) <= 0);
1607 KKASSERT(hammer_btree_cmp(cursor->left_bound, &elm->leaf.base) <= 0);
1608 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm->leaf.base) > 0);
1609 KKASSERT(hammer_btree_cmp(cursor->right_bound, &elm[1].leaf.base) > 0);
1610
427e5fc6
MD
1611 /*
1612 * If we are at the root of the tree, create a new root node with
1613 * 1 element and split normally. Avoid making major modifications
1614 * until we know the whole operation will work.
1615 */
8cd0a023 1616 if (ondisk->parent == 0) {
36f82b23 1617 parent = hammer_alloc_btree(cursor->trans, &error);
427e5fc6 1618 if (parent == NULL)
6a37e7e4 1619 goto done;
8cd0a023 1620 hammer_lock_ex(&parent->lock);
36f82b23 1621 hammer_modify_node_noundo(cursor->trans, parent);
8cd0a023
MD
1622 ondisk = parent->ondisk;
1623 ondisk->count = 1;
1624 ondisk->parent = 0;
1625 ondisk->type = HAMMER_BTREE_TYPE_INTERNAL;
47197d71 1626 ondisk->elms[0].base = hmp->root_btree_beg;
fe7678ee 1627 ondisk->elms[0].base.btype = leaf->ondisk->type;
8cd0a023 1628 ondisk->elms[0].internal.subtree_offset = leaf->node_offset;
47197d71 1629 ondisk->elms[1].base = hmp->root_btree_end;
fe7678ee 1630 /* ondisk->elms[1].base.btype = not used */
10a5d1ba 1631 hammer_modify_node_done(parent);
427e5fc6 1632 made_root = 1;
8cd0a023 1633 parent_index = 0; /* insertion point in parent */
427e5fc6
MD
1634 } else {
1635 made_root = 0;
8cd0a023
MD
1636 parent = cursor->parent;
1637 parent_index = cursor->parent_index;
427e5fc6 1638 }
427e5fc6
MD
1639
1640 /*
1641 * Split leaf into new_leaf at the split point. Select a separator
1642 * value in-between the two leafs but with a bent towards the right
1643 * leaf since comparisons use an 'elm >= separator' inequality.
1644 *
1645 * L L L L L L L L
1646 *
1647 * x x P x x
1648 * s S S s
1649 * / \
1650 * L L L L L L L L
1651 */
36f82b23 1652 new_leaf = hammer_alloc_btree(cursor->trans, &error);
427e5fc6 1653 if (new_leaf == NULL) {
8cd0a023
MD
1654 if (made_root) {
1655 hammer_unlock(&parent->lock);
36f82b23 1656 hammer_delete_node(cursor->trans, parent);
8cd0a023
MD
1657 hammer_rel_node(parent);
1658 }
6a37e7e4 1659 goto done;
427e5fc6 1660 }
8cd0a023 1661 hammer_lock_ex(&new_leaf->lock);
427e5fc6
MD
1662
1663 /*
36f82b23
MD
1664 * Create the new node and copy the leaf elements from the split
1665 * point on to the new node.
427e5fc6 1666 */
36f82b23
MD
1667 hammer_modify_node_all(cursor->trans, leaf);
1668 hammer_modify_node_noundo(cursor->trans, new_leaf);
8cd0a023
MD
1669 ondisk = leaf->ondisk;
1670 elm = &ondisk->elms[split];
1671 bcopy(elm, &new_leaf->ondisk->elms[0], (ondisk->count - split) * esize);
1672 new_leaf->ondisk->count = ondisk->count - split;
1673 new_leaf->ondisk->parent = parent->node_offset;
1674 new_leaf->ondisk->type = HAMMER_BTREE_TYPE_LEAF;
1675 KKASSERT(ondisk->type == new_leaf->ondisk->type);
10a5d1ba 1676 hammer_modify_node_done(new_leaf);
b3bad96f 1677 hammer_cursor_split_node(leaf, new_leaf, split);
427e5fc6
MD
1678
1679 /*
8cd0a023
MD
1680 * Cleanup the original node. Because this is a leaf node and
1681 * leaf nodes do not have a right-hand boundary, there
c0ade690
MD
1682 * aren't any special edge cases to clean up. We just fixup the
1683 * count.
427e5fc6 1684 */
c0ade690 1685 ondisk->count = split;
427e5fc6
MD
1686
1687 /*
1688 * Insert the separator into the parent, fixup the parent's
1689 * reference to the original node, and reference the new node.
1690 * The separator is P.
1691 *
1692 * Remember that base.count does not include the right-hand boundary.
1693 * We are copying parent_index+1 to parent_index+2, not +0 to +1.
1694 */
36f82b23 1695 hammer_modify_node_all(cursor->trans, parent);
8cd0a023 1696 ondisk = parent->ondisk;
36f82b23 1697 KKASSERT(split != 0);
d26d0ae9 1698 KKASSERT(ondisk->count != HAMMER_BTREE_INT_ELMS);
8cd0a023 1699 parent_elm = &ondisk->elms[parent_index+1];
d26d0ae9
MD
1700 bcopy(parent_elm, parent_elm + 1,
1701 (ondisk->count - parent_index) * esize);
eaeff70d 1702
47197d71 1703 hammer_make_separator(&elm[-1].base, &elm[0].base, &parent_elm->base);
fe7678ee 1704 parent_elm->internal.base.btype = new_leaf->ondisk->type;
8cd0a023 1705 parent_elm->internal.subtree_offset = new_leaf->node_offset;
b3deaf57 1706 mid_boundary = &parent_elm->base;
76376933 1707 ++ondisk->count;
10a5d1ba 1708 hammer_modify_node_done(parent);
b3bad96f 1709 hammer_cursor_inserted_element(parent, parent_index + 1);
427e5fc6 1710
fe7678ee 1711 /*
47197d71 1712 * The filesystem's root B-Tree pointer may have to be updated.
427e5fc6
MD
1713 */
1714 if (made_root) {
47197d71
MD
1715 hammer_volume_t volume;
1716
1717 volume = hammer_get_root_volume(hmp, &error);
1718 KKASSERT(error == 0);
1719
e8599db1
MD
1720 hammer_modify_volume_field(cursor->trans, volume,
1721 vol0_btree_root);
47197d71 1722 volume->ondisk->vol0_btree_root = parent->node_offset;
10a5d1ba 1723 hammer_modify_volume_done(volume);
8cd0a023
MD
1724 leaf->ondisk->parent = parent->node_offset;
1725 if (cursor->parent) {
1726 hammer_unlock(&cursor->parent->lock);
1727 hammer_rel_node(cursor->parent);
1728 }
1729 cursor->parent = parent; /* lock'd and ref'd */
47197d71 1730 hammer_rel_volume(volume, 0);
427e5fc6 1731 }
10a5d1ba 1732 hammer_modify_node_done(leaf);
8cd0a023 1733
427e5fc6
MD
1734 /*
1735 * Ok, now adjust the cursor depending on which element the original
1736 * index was pointing at. If we are >= the split point the push node
1737 * is now in the new node.
1738 *
b3deaf57
MD
1739 * NOTE: If we are at the split point itself we need to select the
1740 * old or new node based on where key_beg's insertion point will be.
1741 * If we pick the wrong side the inserted element will wind up in
1742 * the wrong leaf node and outside that node's bounds.
427e5fc6 1743 */
b3deaf57
MD
1744 if (cursor->index > split ||
1745 (cursor->index == split &&
1746 hammer_btree_cmp(&cursor->key_beg, mid_boundary) >= 0)) {
8cd0a023 1747 cursor->parent_index = parent_index + 1;
427e5fc6 1748 cursor->index -= split;
8cd0a023
MD
1749 hammer_unlock(&cursor->node->lock);
1750 hammer_rel_node(cursor->node);
1751 cursor->node = new_leaf;
1752 } else {
1753 cursor->parent_index = parent_index;
1754 hammer_unlock(&new_leaf->lock);
1755 hammer_rel_node(new_leaf);
427e5fc6 1756 }
76376933
MD
1757
1758 /*
1759 * Fixup left and right bounds
1760 */
1761 parent_elm = &parent->ondisk->elms[cursor->parent_index];
fbc6e32a
MD
1762 cursor->left_bound = &parent_elm[0].internal.base;
1763 cursor->right_bound = &parent_elm[1].internal.base;
eaeff70d
MD
1764
1765 /*
47197d71 1766 * Assert that the bounds are correct.
eaeff70d 1767 */
b3deaf57
MD
1768 KKASSERT(hammer_btree_cmp(cursor->left_bound,
1769 &cursor->node->ondisk->elms[0].leaf.base) <= 0);
1770 KKASSERT(hammer_btree_cmp(cursor->right_bound,
47197d71 1771 &cursor->node->ondisk->elms[cursor->node->ondisk->count-1].leaf.base) > 0);
36f82b23
MD
1772 KKASSERT(hammer_btree_cmp(cursor->left_bound, &cursor->key_beg) <= 0);
1773 KKASSERT(hammer_btree_cmp(cursor->right_bound, &cursor->key_beg) > 0);
76376933 1774
6a37e7e4
MD
1775done:
1776 hammer_cursor_downgrade(cursor);
1777 return (error);
427e5fc6
MD
1778}
1779
32c90105
MD
1780/*
1781 * Recursively correct the right-hand boundary's create_tid to (tid) as
1782 * long as the rest of the key matches. We have to recurse upward in
1783 * the tree as well as down the left side of each parent's right node.
1784 *
1785 * Return EDEADLK if we were only partially successful, forcing the caller
1786 * to try again. The original cursor is not modified. This routine can
1787 * also fail with EDEADLK if it is forced to throw away a portion of its
1788 * record history.
1789 *
1790 * The caller must pass a downgraded cursor to us (otherwise we can't dup it).
1791 */
1792struct hammer_rhb {
1793 TAILQ_ENTRY(hammer_rhb) entry;
1794 hammer_node_t node;
1795 int index;
1796};
1797
1798TAILQ_HEAD(hammer_rhb_list, hammer_rhb);
1799
1800int
1801hammer_btree_correct_rhb(hammer_cursor_t cursor, hammer_tid_t tid)
1802{
1803 struct hammer_rhb_list rhb_list;
1804 hammer_base_elm_t elm;
1805 hammer_node_t orig_node;
1806 struct hammer_rhb *rhb;
1807 int orig_index;
1808 int error;
1809
1810 TAILQ_INIT(&rhb_list);
1811
1812 /*
1813 * Save our position so we can restore it on return. This also
1814 * gives us a stable 'elm'.
1815 */
1816 orig_node = cursor->node;
1817 hammer_ref_node(orig_node);
1818 hammer_lock_sh(&orig_node->lock);
1819 orig_index = cursor->index;
1820 elm = &orig_node->ondisk->elms[orig_index].base;
1821
1822 /*
1823 * Now build a list of parents going up, allocating a rhb
1824 * structure for each one.
1825 */
1826 while (cursor->parent) {
1827 /*
1828 * Stop if we no longer have any right-bounds to fix up
1829 */
1830 if (elm->obj_id != cursor->right_bound->obj_id ||
1831 elm->rec_type != cursor->right_bound->rec_type ||
1832 elm->key != cursor->right_bound->key) {
1833 break;
1834 }
1835
1836 /*
1837 * Stop if the right-hand bound's create_tid does not
47197d71 1838 * need to be corrected.
32c90105
MD
1839 */
1840 if (cursor->right_bound->create_tid >= tid)
1841 break;
1842
32c90105
MD
1843 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1844 rhb->node = cursor->parent;
1845 rhb->index = cursor->parent_index;
1846 hammer_ref_node(rhb->node);
1847 hammer_lock_sh(&rhb->node->lock);
1848 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1849
1850 hammer_cursor_up(cursor);
1851 }
1852
1853 /*
1854 * now safely adjust the right hand bound for each rhb. This may
1855 * also require taking the right side of the tree and iterating down
1856 * ITS left side.
1857 */
1858 error = 0;
1859 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1860 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
32c90105
MD
1861 if (error)
1862 break;
1863 TAILQ_REMOVE(&rhb_list, rhb, entry);
1864 hammer_unlock(&rhb->node->lock);
1865 hammer_rel_node(rhb->node);
1866 kfree(rhb, M_HAMMER);
1867
1868 switch (cursor->node->ondisk->type) {
1869 case HAMMER_BTREE_TYPE_INTERNAL:
1870 /*
1871 * Right-boundary for parent at internal node
1872 * is one element to the right of the element whos
1873 * right boundary needs adjusting. We must then
1874 * traverse down the left side correcting any left
1875 * bounds (which may now be too far to the left).
1876 */
1877 ++cursor->index;
1878 error = hammer_btree_correct_lhb(cursor, tid);
1879 break;
32c90105
MD
1880 default:
1881 panic("hammer_btree_correct_rhb(): Bad node type");
1882 error = EINVAL;
1883 break;
1884 }
1885 }
1886
1887 /*
1888 * Cleanup
1889 */
1890 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1891 TAILQ_REMOVE(&rhb_list, rhb, entry);
1892 hammer_unlock(&rhb->node->lock);
1893 hammer_rel_node(rhb->node);
1894 kfree(rhb, M_HAMMER);
1895 }
1896 error = hammer_cursor_seek(cursor, orig_node, orig_index);
1897 hammer_unlock(&orig_node->lock);
1898 hammer_rel_node(orig_node);
1899 return (error);
1900}
1901
1902/*
1903 * Similar to rhb (in fact, rhb calls lhb), but corrects the left hand
1904 * bound going downward starting at the current cursor position.
1905 *
1906 * This function does not restore the cursor after use.
1907 */
1908int
1909hammer_btree_correct_lhb(hammer_cursor_t cursor, hammer_tid_t tid)
1910{
1911 struct hammer_rhb_list rhb_list;
1912 hammer_base_elm_t elm;
1913 hammer_base_elm_t cmp;
1914 struct hammer_rhb *rhb;
1915 int error;
1916
1917 TAILQ_INIT(&rhb_list);
1918
1919 cmp = &cursor->node->ondisk->elms[cursor->index].base;
1920
1921 /*
1922 * Record the node and traverse down the left-hand side for all
1923 * matching records needing a boundary correction.
1924 */
1925 error = 0;
1926 for (;;) {
1927 rhb = kmalloc(sizeof(*rhb), M_HAMMER, M_WAITOK|M_ZERO);
1928 rhb->node = cursor->node;
1929 rhb->index = cursor->index;
1930 hammer_ref_node(rhb->node);
1931 hammer_lock_sh(&rhb->node->lock);
1932 TAILQ_INSERT_HEAD(&rhb_list, rhb, entry);
1933
1934 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
1935 /*
1936 * Nothing to traverse down if we are at the right
1937 * boundary of an internal node.
1938 */
1939 if (cursor->index == cursor->node->ondisk->count)
1940 break;
1941 } else {
1942 elm = &cursor->node->ondisk->elms[cursor->index].base;
1943 if (elm->btype == HAMMER_BTREE_TYPE_RECORD)
1944 break;
47197d71 1945 panic("Illegal leaf record type %02x", elm->btype);
32c90105
MD
1946 }
1947 error = hammer_cursor_down(cursor);
1948 if (error)
1949 break;
1950
1951 elm = &cursor->node->ondisk->elms[cursor->index].base;
1952 if (elm->obj_id != cmp->obj_id ||
1953 elm->rec_type != cmp->rec_type ||
1954 elm->key != cmp->key) {
1955 break;
1956 }
1957 if (elm->create_tid >= tid)
1958 break;
1959
1960 }
1961
1962 /*
1963 * Now we can safely adjust the left-hand boundary from the bottom-up.
1964 * The last element we remove from the list is the caller's right hand
1965 * boundary, which must also be adjusted.
1966 */
1967 while (error == 0 && (rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1968 error = hammer_cursor_seek(cursor, rhb->node, rhb->index);
1969 if (error)
1970 break;
1971 TAILQ_REMOVE(&rhb_list, rhb, entry);
1972 hammer_unlock(&rhb->node->lock);
1973 hammer_rel_node(rhb->node);
1974 kfree(rhb, M_HAMMER);
1975
1976 elm = &cursor->node->ondisk->elms[cursor->index].base;
1977 if (cursor->node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
36f82b23 1978 hammer_modify_node(cursor->trans, cursor->node,
19619882
MD
1979 &elm->create_tid,
1980 sizeof(elm->create_tid));
32c90105 1981 elm->create_tid = tid;
10a5d1ba 1982 hammer_modify_node_done(cursor->node);
32c90105
MD
1983 } else {
1984 panic("hammer_btree_correct_lhb(): Bad element type");
1985 }
1986 }
1987
1988 /*
1989 * Cleanup
1990 */
1991 while ((rhb = TAILQ_FIRST(&rhb_list)) != NULL) {
1992 TAILQ_REMOVE(&rhb_list, rhb, entry);
1993 hammer_unlock(&rhb->node->lock);
1994 hammer_rel_node(rhb->node);
1995 kfree(rhb, M_HAMMER);
1996 }
1997 return (error);
1998}
1999
427e5fc6 2000/*
f36a9737
MD
2001 * Attempt to remove the locked, empty or want-to-be-empty B-Tree node at
2002 * (cursor->node). Returns 0 on success, EDEADLK if we could not complete
2003 * the operation due to a deadlock, or some other error.
8cd0a023 2004 *
f36a9737
MD
2005 * This routine is always called with an empty, locked leaf but may recurse
2006 * into want-to-be-empty parents as part of its operation.
b3deaf57 2007 *
c82af904
MD
2008 * It should also be noted that when removing empty leaves we must be sure
2009 * to test and update mirror_tid because another thread may have deadlocked
5fa5c92f 2010 * against us (or someone) trying to propagate it up and cannot retry once
c82af904
MD
2011 * the node has been deleted.
2012 *
f36a9737
MD
2013 * On return the cursor may end up pointing to an internal node, suitable
2014 * for further iteration but not for an immediate insertion or deletion.
8cd0a023 2015 */
f36a9737 2016static int
46fe7ae1 2017btree_remove(hammer_cursor_t cursor)
8cd0a023
MD
2018{
2019 hammer_node_ondisk_t ondisk;
195c19a1 2020 hammer_btree_elm_t elm;
195c19a1 2021 hammer_node_t node;
8cd0a023 2022 hammer_node_t parent;
fe7678ee 2023 const int esize = sizeof(*elm);
8cd0a023 2024 int error;
8cd0a023 2025
fe7678ee
MD
2026 node = cursor->node;
2027
47197d71
MD
2028 /*
2029 * When deleting the root of the filesystem convert it to
2030 * an empty leaf node. Internal nodes cannot be empty.
2031 */
c82af904
MD
2032 ondisk = node->ondisk;
2033 if (ondisk->parent == 0) {
f36a9737 2034 KKASSERT(cursor->parent == NULL);
36f82b23 2035 hammer_modify_node_all(cursor->trans, node);
c82af904 2036 KKASSERT(ondisk == node->ondisk);
195c19a1
MD
2037 ondisk->type = HAMMER_BTREE_TYPE_LEAF;
2038 ondisk->count = 0;
10a5d1ba 2039 hammer_modify_node_done(node);
b3deaf57 2040 cursor->index = 0;
47197d71 2041 return(0);
8cd0a023
MD
2042 }
2043
c82af904 2044 parent = cursor->parent;
b3bad96f 2045 hammer_cursor_removed_node(node, parent, cursor->parent_index);
c82af904 2046
8cd0a023 2047 /*
f36a9737
MD
2048 * Attempt to remove the parent's reference to the child. If the
2049 * parent would become empty we have to recurse. If we fail we
2050 * leave the parent pointing to an empty leaf node.
8cd0a023 2051 */
f36a9737
MD
2052 if (parent->ondisk->count == 1) {
2053 /*
2054 * This special cursor_up_locked() call leaves the original
2055 * node exclusively locked and referenced, leaves the
2056 * original parent locked (as the new node), and locks the
2057 * new parent. It can return EDEADLK.
2058 */
2059 error = hammer_cursor_up_locked(cursor);
2060 if (error == 0) {
2061 error = btree_remove(cursor);
2062 if (error == 0) {
2063 hammer_modify_node_all(cursor->trans, node);
2064 ondisk = node->ondisk;
2065 ondisk->type = HAMMER_BTREE_TYPE_DELETED;
2066 ondisk->count = 0;
2067 hammer_modify_node_done(node);
2068 hammer_flush_node(node);
2069 hammer_delete_node(cursor->trans, node);
2070 } else {
2071 kprintf("Warning: BTREE_REMOVE: Defering "
2072 "parent removal1 @ %016llx, skipping\n",
2073 node->node_offset);
2074 }
2075 hammer_unlock(&node->lock);
2076 hammer_rel_node(node);
2077 } else {
2078 kprintf("Warning: BTREE_REMOVE: Defering parent "
2079 "removal2 @ %016llx, skipping\n",
2080 node->node_offset);
2081 }
2082 } else {
2083 KKASSERT(parent->ondisk->count > 1);
6a37e7e4 2084
f36a9737
MD
2085 /*
2086 * Delete the subtree reference in the parent
2087 */
2088 hammer_modify_node_all(cursor->trans, parent);
2089 ondisk = parent->ondisk;
2090 KKASSERT(ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
c82af904 2091
f36a9737
MD
2092 elm = &ondisk->elms[cursor->parent_index];
2093 KKASSERT(elm->internal.subtree_offset == node->node_offset);
2094 KKASSERT(ondisk->count > 0);
2095 bcopy(&elm[1], &elm[0],
2096 (ondisk->count - cursor->parent_index) * esize);
2097 --ondisk->count;
10a5d1ba 2098 hammer_modify_node_done(parent);
f36a9737
MD
2099 hammer_flush_node(node);
2100 hammer_delete_node(cursor->trans, node);
6a37e7e4 2101
f36a9737
MD
2102 /*
2103 * cursor->node is invalid, cursor up to make the cursor
2104 * valid again.
2105 */
2106 error = hammer_cursor_up(cursor);
6a37e7e4 2107 }
f36a9737 2108 return (error);
6a37e7e4
MD
2109}
2110
602c6cb8
MD
2111/*
2112 * Propagate cursor->trans->tid up the B-Tree starting at the current
2113 * cursor position using pseudofs info gleaned from the passed inode.
2114 *
2115 * The passed inode has no relationship to the cursor position other
2116 * then being in the same pseudofs as the insertion or deletion we
2117 * are propagating the mirror_tid for.
2118 */
2119void
2120hammer_btree_do_propagation(hammer_cursor_t cursor, hammer_inode_t ip,
2121 hammer_btree_leaf_elm_t leaf)
2122{
2123 hammer_pseudofs_inmem_t pfsm;
2124 int error;
2125
2126 /*
2127 * We only propagate the mirror_tid up if we are in master or slave
2128 * mode. We do not bother if we are in no-mirror mode.
2129 */
2130 pfsm = ip->pfsm;
2131 KKASSERT(pfsm != NULL);
2132 if (pfsm->pfsd.master_id < 0 &&
2133 (pfsm->pfsd.mirror_flags & HAMMER_PFSD_SLAVE) == 0) {
2134 return;
2135 }
2136
602c6cb8
MD
2137 error = hammer_btree_mirror_propagate(cursor->trans,
2138 cursor->parent, cursor->parent_index,
2139 cursor->node->ondisk->mirror_tid);
2140 /* XXX */
2141}
2142
2143
c82af904
MD
2144/*
2145 * Propagate a mirror TID update upwards through the B-Tree to the root.
2146 *
2147 * A locked internal node must be passed in. The node will remain locked
2148 * on return.
2149 *
2150 * This function syncs mirror_tid at the specified internal node's element,
2151 * adjusts the node's aggregation mirror_tid, and then recurses upwards.
2152 */
602c6cb8 2153static int
c82af904
MD
2154hammer_btree_mirror_propagate(hammer_transaction_t trans, hammer_node_t node,
2155 int index, hammer_tid_t mirror_tid)
2156{
2157 hammer_btree_internal_elm_t elm;
2158 hammer_node_t parent;
2159 int parent_index;
2160 int error;
2161
2162 KKASSERT (node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL);
2163
2164 /*
2165 * Adjust the node's element
2166 */
2167 elm = &node->ondisk->elms[index].internal;
2168 if (elm->mirror_tid >= mirror_tid)
2169 return(0);
2170 hammer_modify_node(trans, node, &elm->mirror_tid,
2171 sizeof(elm->mirror_tid));
2172 elm->mirror_tid = mirror_tid;
2173 hammer_modify_node_done(node);
2174
2175 /*
602c6cb8 2176 * Adjust the node's mirror_tid aggregator
c82af904
MD
2177 */
2178 if (node->ondisk->mirror_tid >= mirror_tid)
2179 return(0);
2180 hammer_modify_node_field(trans, node, mirror_tid);
2181 node->ondisk->mirror_tid = mirror_tid;
2182 hammer_modify_node_done(node);
2183
2184 error = 0;
5fa5c92f 2185 if (node->ondisk->parent) {
c82af904
MD
2186 parent = hammer_btree_get_parent(node, &parent_index,
2187 &error, 1);
2188 if (parent) {
2189 hammer_btree_mirror_propagate(trans, parent,
2190 parent_index, mirror_tid);
2191 hammer_unlock(&parent->lock);
2192 hammer_rel_node(parent);
2193 }
2194 }
2195 return(error);
2196}
2197
2198hammer_node_t
2199hammer_btree_get_parent(hammer_node_t node, int *parent_indexp, int *errorp,
2200 int try_exclusive)
2201{
2202 hammer_node_t parent;
2203 hammer_btree_elm_t elm;
2204 int i;
2205
2206 /*
2207 * Get the node
2208 */
2209 parent = hammer_get_node(node->hmp, node->ondisk->parent, 0, errorp);
2210 if (*errorp) {
2211 KKASSERT(parent == NULL);
2212 return(NULL);
2213 }
2214 KKASSERT ((parent->flags & HAMMER_NODE_DELETED) == 0);
2215
2216 /*
2217 * Lock the node
2218 */
2219 if (try_exclusive) {
2220 if (hammer_lock_ex_try(&parent->lock)) {
2221 hammer_rel_node(parent);
2222 *errorp = EDEADLK;
2223 return(NULL);
2224 }
2225 } else {
2226 hammer_lock_sh(&parent->lock);
2227 }
2228
2229 /*
2230 * Figure out which element in the parent is pointing to the
2231 * child.
2232 */
2233 if (node->ondisk->count) {
2234 i = hammer_btree_search_node(&node->ondisk->elms[0].base,
2235 parent->ondisk);
2236 } else {
2237 i = 0;
2238 }
2239 while (i < parent->ondisk->count) {
2240 elm = &parent->ondisk->elms[i];
2241 if (elm->internal.subtree_offset == node->node_offset)
2242 break;
2243 ++i;
2244 }
2245 if (i == parent->ondisk->count) {
2246 hammer_unlock(&parent->lock);
2247 panic("Bad B-Tree link: parent %p node %p\n", parent, node);
2248 }
2249 *parent_indexp = i;
2250 KKASSERT(*errorp == 0);
2251 return(parent);
2252}
2253
7f7c1f84 2254/*
fe7678ee
MD
2255 * The element (elm) has been moved to a new internal node (node).
2256 *
2257 * If the element represents a pointer to an internal node that node's
2258 * parent must be adjusted to the element's new location.
2259 *
6a37e7e4 2260 * XXX deadlock potential here with our exclusive locks
7f7c1f84 2261 */
7f7c1f84 2262int
36f82b23
MD
2263btree_set_parent(hammer_transaction_t trans, hammer_node_t node,
2264 hammer_btree_elm_t elm)
7f7c1f84 2265{
7f7c1f84
MD
2266 hammer_node_t child;
2267 int error;
2268
2269 error = 0;
2270
fe7678ee 2271 switch(elm->base.btype) {
7f7c1f84 2272 case HAMMER_BTREE_TYPE_INTERNAL:
fe7678ee 2273 case HAMMER_BTREE_TYPE_LEAF:
19619882
MD
2274 child = hammer_get_node(node->hmp, elm->internal.subtree_offset,
2275 0, &error);
7f7c1f84 2276 if (error == 0) {
c9b9e29d 2277 hammer_modify_node_field(trans, child, parent);
7f7c1f84 2278 child->ondisk->parent = node->node_offset;
10a5d1ba 2279 hammer_modify_node_done(child);
7f7c1f84
MD
2280 hammer_rel_node(child);
2281 }
2282 break;
7f7c1f84 2283 default:
fe7678ee 2284 break;
7f7c1f84
MD
2285 }
2286 return(error);
2287}
2288
b33e2cc0
MD
2289/*
2290 * Exclusively lock all the children of node. This is used by the split
2291 * code to prevent anyone from accessing the children of a cursor node
2292 * while we fix-up its parent offset.
2293 *
2294 * If we don't lock the children we can really mess up cursors which block
2295 * trying to cursor-up into our node.
2296 *
b33e2cc0
MD
2297 * On failure EDEADLK (or some other error) is returned. If a deadlock
2298 * error is returned the cursor is adjusted to block on termination.
2299 */
2300int
2301hammer_btree_lock_children(hammer_cursor_t cursor,
2302 struct hammer_node_locklist **locklistp)
2303{
2304 hammer_node_t node;
2305 hammer_node_locklist_t item;
2306 hammer_node_ondisk_t ondisk;
2307 hammer_btree_elm_t elm;
b33e2cc0
MD
2308 hammer_node_t child;
2309 int error;
2310 int i;
2311
2312 node = cursor->node;
2313 ondisk = node->ondisk;
2314 error = 0;
7bc5b8c2
MD
2315
2316 /*
2317 * We really do not want to block on I/O with exclusive locks held,
2318 * pre-get the children before trying to lock the mess.
2319 */
2320 for (i = 0; i < ondisk->count; ++i) {
cb51be26 2321 ++hammer_stats_btree_elements;
7bc5b8c2
MD
2322 elm = &ondisk->elms[i];
2323 if (elm->base.btype != HAMMER_BTREE_TYPE_LEAF &&
2324 elm->base.btype != HAMMER_BTREE_TYPE_INTERNAL) {
2325 continue;
2326 }
2327 child = hammer_get_node(node->hmp,
2328 elm->internal.subtree_offset,
2329 0, &error);
2330 if (child)
2331 hammer_rel_node(child);
2332 }
2333
2334 /*
2335 * Do it for real
2336 */
b33e2cc0 2337 for (i = 0; error == 0 && i < ondisk->count; ++i) {
cb51be26 2338 ++hammer_stats_btree_elements;
b33e2cc0
MD
2339 elm = &ondisk->elms[i];
2340
b33e2cc0
MD
2341 switch(elm->base.btype) {
2342 case HAMMER_BTREE_TYPE_INTERNAL:
2343 case HAMMER_BTREE_TYPE_LEAF:
f36a9737 2344 KKASSERT(elm->internal.subtree_offset != 0);
40043e7f 2345 child = hammer_get_node(node->hmp,
b33e2cc0 2346 elm->internal.subtree_offset,
19619882 2347 0, &error);
b33e2cc0 2348 break;
b33e2cc0 2349 default:
47197d71 2350 child = NULL;
b33e2cc0
MD
2351 break;
2352 }
2353 if (child) {
2354 if (hammer_lock_ex_try(&child->lock) != 0) {
2355 if (cursor->deadlk_node == NULL) {
1ff9f58e 2356 cursor->deadlk_node = child;
b33e2cc0
MD
2357 hammer_ref_node(cursor->deadlk_node);
2358 }
2359 error = EDEADLK;
a84a197d 2360 hammer_rel_node(child);
b33e2cc0
MD
2361 } else {
2362 item = kmalloc(sizeof(*item),
2363 M_HAMMER, M_WAITOK);
2364 item->next = *locklistp;
2365 item->node = child;
2366 *locklistp = item;
2367 }
2368 }
2369 }
2370 if (error)
2371 hammer_btree_unlock_children(locklistp);
2372 return(error);
2373}
2374
2375
2376/*
2377 * Release previously obtained node locks.
2378 */
2f85fa4d 2379void
b33e2cc0
MD
2380hammer_btree_unlock_children(struct hammer_node_locklist **locklistp)
2381{
2382 hammer_node_locklist_t item;
2383
2384 while ((item = *locklistp) != NULL) {
2385 *locklistp = item->next;
2386 hammer_unlock(&item->node->lock);
2387 hammer_rel_node(item->node);
2388 kfree(item, M_HAMMER);
2389 }
2390}
2391
8cd0a023
MD
2392/************************************************************************
2393 * MISCELLANIOUS SUPPORT *
2394 ************************************************************************/
2395
2396/*
d26d0ae9 2397 * Compare two B-Tree elements, return -N, 0, or +N (e.g. similar to strcmp).
8cd0a023 2398 *
d113fda1 2399 * Note that for this particular function a return value of -1, 0, or +1
9582c7da 2400 * can denote a match if create_tid is otherwise discounted. A create_tid
d5530d22 2401 * of zero is considered to be 'infinity' in comparisons.
d113fda1 2402 *
8cd0a023 2403 * See also hammer_rec_rb_compare() and hammer_rec_cmp() in hammer_object.c.
8cd0a023
MD
2404 */
2405int
2406hammer_btree_cmp(hammer_base_elm_t key1, hammer_base_elm_t key2)
2407{
2f85fa4d
MD
2408 if (key1->localization < key2->localization)
2409 return(-5);
2410 if (key1->localization > key2->localization)
2411 return(5);
2412
d26d0ae9
MD
2413 if (key1->obj_id < key2->obj_id)
2414 return(-4);
2415 if (key1->obj_id > key2->obj_id)
2416 return(4);
8cd0a023 2417
d26d0ae9
MD
2418 if (key1->rec_type < key2->rec_type)
2419 return(-3);
2420 if (key1->rec_type > key2->rec_type)
2421 return(3);
8cd0a023 2422
8cd0a023
MD
2423 if (key1->key < key2->key)
2424 return(-2);
2425 if (key1->key > key2->key)
2426 return(2);
d113fda1 2427
d5530d22 2428 /*
9582c7da
MD
2429 * A create_tid of zero indicates a record which is undeletable
2430 * and must be considered to have a value of positive infinity.
d5530d22 2431 */
9582c7da
MD
2432 if (key1->create_tid == 0) {
2433 if (key2->create_tid == 0)
d5530d22
MD
2434 return(0);
2435 return(1);
2436 }
9582c7da 2437 if (key2->create_tid == 0)
d5530d22 2438 return(-1);
9582c7da 2439 if (key1->create_tid < key2->create_tid)
d113fda1 2440 return(-1);
9582c7da 2441 if (key1->create_tid > key2->create_tid)
d113fda1 2442 return(1);
8cd0a023
MD
2443 return(0);
2444}
2445
c0ade690 2446/*
d5530d22
MD
2447 * Test a timestamp against an element to determine whether the
2448 * element is visible. A timestamp of 0 means 'infinity'.
c0ade690
MD
2449 */
2450int
d5530d22 2451hammer_btree_chkts(hammer_tid_t asof, hammer_base_elm_t base)
c0ade690 2452{
d5530d22
MD
2453 if (asof == 0) {
2454 if (base->delete_tid)
2455 return(1);
2456 return(0);
2457 }
2458 if (asof < base->create_tid)
d26d0ae9 2459 return(-1);
d5530d22 2460 if (base->delete_tid && asof >= base->delete_tid)
d26d0ae9 2461 return(1);
c0ade690
MD
2462 return(0);
2463}
2464
8cd0a023
MD
2465/*
2466 * Create a separator half way inbetween key1 and key2. For fields just
d5530d22
MD
2467 * one unit apart, the separator will match key2. key1 is on the left-hand
2468 * side and key2 is on the right-hand side.
8cd0a023 2469 *
9391cded 2470 * key2 must be >= the separator. It is ok for the separator to match key2.
36f82b23 2471 *
9391cded
MD
2472 * NOTE: Even if key1 does not match key2, the separator may wind up matching
2473 * key2.
2474 *
2475 * NOTE: It might be beneficial to just scrap this whole mess and just
2476 * set the separator to key2.
8cd0a023
MD
2477 */
2478#define MAKE_SEPARATOR(key1, key2, dest, field) \
2479 dest->field = key1->field + ((key2->field - key1->field + 1) >> 1);
2480
2481static void
2482hammer_make_separator(hammer_base_elm_t key1, hammer_base_elm_t key2,
2483 hammer_base_elm_t dest)
2484{
2485 bzero(dest, sizeof(*dest));
d5530d22 2486
9391cded
MD
2487 dest->rec_type = key2->rec_type;
2488 dest->key = key2->key;
2f85fa4d 2489 dest->obj_id = key2->obj_id;
9391cded
MD
2490 dest->create_tid = key2->create_tid;
2491
2f85fa4d
MD
2492 MAKE_SEPARATOR(key1, key2, dest, localization);
2493 if (key1->localization == key2->localization) {
2494 MAKE_SEPARATOR(key1, key2, dest, obj_id);
2495 if (key1->obj_id == key2->obj_id) {
2496 MAKE_SEPARATOR(key1, key2, dest, rec_type);
2497 if (key1->rec_type == key2->rec_type) {
2498 MAKE_SEPARATOR(key1, key2, dest, key);
2499 /*
2500 * Don't bother creating a separator for
2501 * create_tid, which also conveniently avoids
2502 * having to handle the create_tid == 0
2503 * (infinity) case. Just leave create_tid
2504 * set to key2.
2505 *
2506 * Worst case, dest matches key2 exactly,
2507 * which is acceptable.
2508 */
2509 }
d5530d22 2510 }
d113fda1 2511 }
8cd0a023
MD
2512}
2513
2514#undef MAKE_SEPARATOR
2515
2516/*
2517 * Return whether a generic internal or leaf node is full
2518 */
2519static int
2520btree_node_is_full(hammer_node_ondisk_t node)
2521{
2522 switch(node->type) {
2523 case HAMMER_BTREE_TYPE_INTERNAL:
2524 if (node->count == HAMMER_BTREE_INT_ELMS)
2525 return(1);
2526 break;
2527 case HAMMER_BTREE_TYPE_LEAF:
2528 if (node->count == HAMMER_BTREE_LEAF_ELMS)
2529 return(1);
2530 break;
2531 default:
2532 panic("illegal btree subtype");
2533 }
2534 return(0);
2535}
9944ae54 2536
8cd0a023
MD
2537#if 0
2538static int
2539btree_max_elements(u_int8_t type)
2540{
2541 if (type == HAMMER_BTREE_TYPE_LEAF)
2542 return(HAMMER_BTREE_LEAF_ELMS);
2543 if (type == HAMMER_BTREE_TYPE_INTERNAL)
2544 return(HAMMER_BTREE_INT_ELMS);
2545 panic("btree_max_elements: bad type %d\n", type);
2546}
2547#endif
2548
c0ade690
MD
2549void
2550hammer_print_btree_node(hammer_node_ondisk_t ondisk)
2551{
2552 hammer_btree_elm_t elm;
2553 int i;
2554
47197d71 2555 kprintf("node %p count=%d parent=%016llx type=%c\n",
c0ade690
MD
2556 ondisk, ondisk->count, ondisk->parent, ondisk->type);
2557
2558 /*
2559 * Dump both boundary elements if an internal node
2560 */
2561 if (ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
2562 for (i = 0; i <= ondisk->count; ++i) {
2563 elm = &ondisk->elms[i];
2564 hammer_print_btree_elm(elm, ondisk->type, i);
2565 }
2566 } else {
2567 for (i = 0; i < ondisk->count; ++i) {
2568 elm = &ondisk->elms[i];
2569 hammer_print_btree_elm(elm, ondisk->type, i);
2570 }
2571 }
2572}
2573
2574void
2575hammer_print_btree_elm(hammer_btree_elm_t elm, u_int8_t type, int i)
2576{
2577 kprintf(" %2d", i);
7dc57964 2578 kprintf("\tobj_id = %016llx\n", elm->base.obj_id);
c0ade690
MD
2579 kprintf("\tkey = %016llx\n", elm->base.key);
2580 kprintf("\tcreate_tid = %016llx\n", elm->base.create_tid);
2581 kprintf("\tdelete_tid = %016llx\n", elm->base.delete_tid);
2582 kprintf("\trec_type = %04x\n", elm->base.rec_type);
2583 kprintf("\tobj_type = %02x\n", elm->base.obj_type);
fe7678ee
MD
2584 kprintf("\tbtype = %02x (%c)\n",
2585 elm->base.btype,
2586 (elm->base.btype ? elm->base.btype : '?'));
2f85fa4d 2587 kprintf("\tlocalization = %02x\n", elm->base.localization);
fe7678ee
MD
2588
2589 switch(type) {
2590 case HAMMER_BTREE_TYPE_INTERNAL:
47197d71 2591 kprintf("\tsubtree_off = %016llx\n",
fe7678ee
MD
2592 elm->internal.subtree_offset);
2593 break;
fe7678ee 2594 case HAMMER_BTREE_TYPE_RECORD:
47197d71 2595 kprintf("\tdata_offset = %016llx\n", elm->leaf.data_offset);
c0ade690
MD
2596 kprintf("\tdata_len = %08x\n", elm->leaf.data_len);
2597 kprintf("\tdata_crc = %08x\n", elm->leaf.data_crc);
fe7678ee 2598 break;
c0ade690
MD
2599 }
2600}