HAMMER - Add live dedup sysctl and support
[dragonfly.git] / sys / vfs / hammer / hammer_object.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_object.c,v 1.97 2008/09/23 22:28:56 dillon Exp $
35  */
36
37 #include "hammer.h"
38
39 static int hammer_mem_lookup(hammer_cursor_t cursor);
40 static void hammer_mem_first(hammer_cursor_t cursor);
41 static int hammer_frontend_trunc_callback(hammer_record_t record,
42                                 void *data __unused);
43 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
45 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
46                                 hammer_btree_leaf_elm_t leaf);
47 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
48                                 hammer_btree_leaf_elm_t leaf);
49
50 struct rec_trunc_info {
51         u_int16_t       rec_type;
52         int64_t         trunc_off;
53 };
54
55 struct hammer_bulk_info {
56         hammer_record_t record;
57         hammer_record_t conflict;
58 };
59
60 /*
61  * Red-black tree support.  Comparison code for insertion.
62  */
63 static int
64 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
65 {
66         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
67                 return(-1);
68         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
69                 return(1);
70
71         if (rec1->leaf.base.key < rec2->leaf.base.key)
72                 return(-1);
73         if (rec1->leaf.base.key > rec2->leaf.base.key)
74                 return(1);
75
76         /*
77          * For search & insertion purposes records deleted by the
78          * frontend or deleted/committed by the backend are silently
79          * ignored.  Otherwise pipelined insertions will get messed
80          * up.
81          *
82          * rec1 is greater then rec2 if rec1 is marked deleted.
83          * rec1 is less then rec2 if rec2 is marked deleted.
84          *
85          * Multiple deleted records may be present, do not return 0
86          * if both are marked deleted.
87          */
88         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
89                            HAMMER_RECF_COMMITTED)) {
90                 return(1);
91         }
92         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
93                            HAMMER_RECF_COMMITTED)) {
94                 return(-1);
95         }
96
97         return(0);
98 }
99
100 /*
101  * Basic record comparison code similar to hammer_btree_cmp().
102  *
103  * obj_id is not compared and may not yet be assigned in the record.
104  */
105 static int
106 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
107 {
108         if (elm->rec_type < rec->leaf.base.rec_type)
109                 return(-3);
110         if (elm->rec_type > rec->leaf.base.rec_type)
111                 return(3);
112
113         if (elm->key < rec->leaf.base.key)
114                 return(-2);
115         if (elm->key > rec->leaf.base.key)
116                 return(2);
117
118         /*
119          * Never match against an item deleted by the frontend
120          * or backend, or committed by the backend.
121          *
122          * elm is less then rec if rec is marked deleted.
123          */
124         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
125                           HAMMER_RECF_COMMITTED)) {
126                 return(-1);
127         }
128         return(0);
129 }
130
131 /*
132  * Ranged scan to locate overlapping record(s).  This is used by
133  * hammer_ip_get_bulk() to locate an overlapping record.  We have
134  * to use a ranged scan because the keys for data records with the
135  * same file base offset can be different due to differing data_len's.
136  *
137  * NOTE: The base file offset of a data record is (key - data_len), not (key).
138  */
139 static int
140 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
141 {
142         struct hammer_bulk_info *info = data;
143         hammer_btree_leaf_elm_t leaf = &info->record->leaf;
144
145         if (rec->leaf.base.rec_type < leaf->base.rec_type)
146                 return(-3);
147         if (rec->leaf.base.rec_type > leaf->base.rec_type)
148                 return(3);
149
150         /*
151          * Overlap compare
152          */
153         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
154                 /* rec_beg >= leaf_end */
155                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
156                         return(2);
157                 /* rec_end <= leaf_beg */
158                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
159                         return(-2);
160         } else {
161                 if (rec->leaf.base.key < leaf->base.key)
162                         return(-2);
163                 if (rec->leaf.base.key > leaf->base.key)
164                         return(2);
165         }
166
167         /*
168          * We have to return 0 at this point, even if DELETED_FE is set,
169          * because returning anything else will cause the scan to ignore
170          * one of the branches when we really want it to check both.
171          */
172         return(0);
173 }
174
175 /*
176  * RB_SCAN comparison code for hammer_mem_first().  The argument order
177  * is reversed so the comparison result has to be negated.  key_beg and
178  * key_end are both range-inclusive.
179  *
180  * Localized deletions are not cached in-memory.
181  */
182 static
183 int
184 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
185 {
186         hammer_cursor_t cursor = data;
187         int r;
188
189         r = hammer_rec_cmp(&cursor->key_beg, rec);
190         if (r > 1)
191                 return(-1);
192         r = hammer_rec_cmp(&cursor->key_end, rec);
193         if (r < -1)
194                 return(1);
195         return(0);
196 }
197
198 /*
199  * This compare function is used when simply looking up key_beg.
200  */
201 static
202 int
203 hammer_rec_find_cmp(hammer_record_t rec, void *data)
204 {
205         hammer_cursor_t cursor = data;
206         int r;
207
208         r = hammer_rec_cmp(&cursor->key_beg, rec);
209         if (r > 1)
210                 return(-1);
211         if (r < -1)
212                 return(1);
213         return(0);
214 }
215
216 /*
217  * Locate blocks within the truncation range.  Partial blocks do not count.
218  */
219 static
220 int
221 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
222 {
223         struct rec_trunc_info *info = data;
224
225         if (rec->leaf.base.rec_type < info->rec_type)
226                 return(-1);
227         if (rec->leaf.base.rec_type > info->rec_type)
228                 return(1);
229
230         switch(rec->leaf.base.rec_type) {
231         case HAMMER_RECTYPE_DB:
232                 /*
233                  * DB record key is not beyond the truncation point, retain.
234                  */
235                 if (rec->leaf.base.key < info->trunc_off)
236                         return(-1);
237                 break;
238         case HAMMER_RECTYPE_DATA:
239                 /*
240                  * DATA record offset start is not beyond the truncation point,
241                  * retain.
242                  */
243                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
244                         return(-1);
245                 break;
246         default:
247                 panic("hammer_rec_trunc_cmp: unexpected record type");
248         }
249
250         /*
251          * The record start is >= the truncation point, return match,
252          * the record should be destroyed.
253          */
254         return(0);
255 }
256
257 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
258
259 /*
260  * Allocate a record for the caller to finish filling in.  The record is
261  * returned referenced.
262  */
263 hammer_record_t
264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
265 {
266         hammer_record_t record;
267         hammer_mount_t hmp;
268
269         hmp = ip->hmp;
270         ++hammer_count_records;
271         record = kmalloc(sizeof(*record), hmp->m_misc,
272                          M_WAITOK | M_ZERO | M_USE_RESERVE);
273         record->flush_state = HAMMER_FST_IDLE;
274         record->ip = ip;
275         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
276         record->leaf.data_len = data_len;
277         hammer_ref(&record->lock);
278
279         if (data_len) {
280                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
281                 record->flags |= HAMMER_RECF_ALLOCDATA;
282                 ++hammer_count_record_datas;
283         }
284
285         return (record);
286 }
287
288 void
289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
290 {
291         while (record->flush_state == HAMMER_FST_FLUSH) {
292                 record->flags |= HAMMER_RECF_WANTED;
293                 tsleep(record, 0, ident, 0);
294         }
295 }
296
297 /*
298  * Called from the backend, hammer_inode.c, after a record has been
299  * flushed to disk.  The record has been exclusively locked by the
300  * caller and interlocked with BE.
301  *
302  * We clean up the state, unlock, and release the record (the record
303  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
304  */
305 void
306 hammer_flush_record_done(hammer_record_t record, int error)
307 {
308         hammer_inode_t target_ip;
309
310         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
311         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
312
313         /*
314          * If an error occured, the backend was unable to sync the
315          * record to its media.  Leave the record intact.
316          */
317         if (error) {
318                 hammer_critical_error(record->ip->hmp, record->ip, error,
319                                       "while flushing record");
320         }
321
322         --record->flush_group->refs;
323         record->flush_group = NULL;
324
325         /*
326          * Adjust the flush state and dependancy based on success or
327          * failure.
328          */
329         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
330                 if ((target_ip = record->target_ip) != NULL) {
331                         TAILQ_REMOVE(&target_ip->target_list, record,
332                                      target_entry);
333                         record->target_ip = NULL;
334                         hammer_test_inode(target_ip);
335                 }
336                 record->flush_state = HAMMER_FST_IDLE;
337         } else {
338                 if (record->target_ip) {
339                         record->flush_state = HAMMER_FST_SETUP;
340                         hammer_test_inode(record->ip);
341                         hammer_test_inode(record->target_ip);
342                 } else {
343                         record->flush_state = HAMMER_FST_IDLE;
344                 }
345         }
346         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
347
348         /*
349          * Cleanup
350          */
351         if (record->flags & HAMMER_RECF_WANTED) {
352                 record->flags &= ~HAMMER_RECF_WANTED;
353                 wakeup(record);
354         }
355         hammer_rel_mem_record(record);
356 }
357
358 /*
359  * Release a memory record.  Records marked for deletion are immediately
360  * removed from the RB-Tree but otherwise left intact until the last ref
361  * goes away.
362  */
363 void
364 hammer_rel_mem_record(struct hammer_record *record)
365 {
366         hammer_mount_t hmp;
367         hammer_reserve_t resv;
368         hammer_inode_t ip;
369         hammer_inode_t target_ip;
370         int diddrop;
371
372         hammer_rel(&record->lock);
373
374         if (hammer_norefs(&record->lock)) {
375                 /*
376                  * Upon release of the last reference wakeup any waiters.
377                  * The record structure may get destroyed so callers will
378                  * loop up and do a relookup.
379                  *
380                  * WARNING!  Record must be removed from RB-TREE before we
381                  * might possibly block.  hammer_test_inode() can block!
382                  */
383                 ip = record->ip;
384                 hmp = ip->hmp;
385
386                 /*
387                  * Upon release of the last reference a record marked deleted
388                  * by the front or backend, or committed by the backend,
389                  * is destroyed.
390                  */
391                 if (record->flags & (HAMMER_RECF_DELETED_FE |
392                                      HAMMER_RECF_DELETED_BE |
393                                      HAMMER_RECF_COMMITTED)) {
394                         KKASSERT(hammer_isactive(&ip->lock) > 0);
395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
396
397                         /*
398                          * target_ip may have zero refs, we have to ref it
399                          * to prevent it from being ripped out from under
400                          * us.
401                          */
402                         if ((target_ip = record->target_ip) != NULL) {
403                                 TAILQ_REMOVE(&target_ip->target_list,
404                                              record, target_entry);
405                                 record->target_ip = NULL;
406                                 hammer_ref(&target_ip->lock);
407                         }
408
409                         /*
410                          * Remove the record from the B-Tree
411                          */
412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
413                                 RB_REMOVE(hammer_rec_rb_tree,
414                                           &record->ip->rec_tree,
415                                           record);
416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
417                                 KKASSERT(ip->rsv_recs > 0);
418                                 if (RB_EMPTY(&record->ip->rec_tree)) {
419                                         record->ip->flags &=
420                                                         ~HAMMER_INODE_XDIRTY;
421                                         record->ip->sync_flags &=
422                                                         ~HAMMER_INODE_XDIRTY;
423                                 }
424                                 diddrop = 1;
425                         } else {
426                                 diddrop = 0;
427                         }
428
429                         /*
430                          * We must wait for any direct-IO to complete before
431                          * we can destroy the record because the bio may
432                          * have a reference to it.
433                          */
434                         if (record->gflags &
435                            (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
436                                 hammer_io_direct_wait(record);
437                         }
438
439                         /*
440                          * Account for the completion after the direct IO
441                          * has completed.
442                          */
443                         if (diddrop) {
444                                 --hmp->rsv_recs;
445                                 --ip->rsv_recs;
446                                 hmp->rsv_databytes -= record->leaf.data_len;
447
448                                 if (RB_EMPTY(&record->ip->rec_tree))
449                                         hammer_test_inode(record->ip);
450                                 if (ip->rsv_recs == hammer_limit_inode_recs - 1)
451                                         wakeup(&ip->rsv_recs);
452                         }
453
454                         /*
455                          * Do this test after removing record from the B-Tree.
456                          */
457                         if (target_ip) {
458                                 hammer_test_inode(target_ip);
459                                 hammer_rel_inode(target_ip, 0);
460                         }
461
462                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
463                                 --hammer_count_record_datas;
464                                 kfree(record->data, hmp->m_misc);
465                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
466                         }
467
468                         /*
469                          * Release the reservation.
470                          *
471                          * If the record was not committed we can theoretically
472                          * undo the reservation.  However, doing so might
473                          * create weird edge cases with the ordering of
474                          * direct writes because the related buffer cache
475                          * elements are per-vnode.  So we don't try.
476                          */
477                         if ((resv = record->resv) != NULL) {
478                                 /* XXX undo leaf.data_offset,leaf.data_len */
479                                 hammer_blockmap_reserve_complete(hmp, resv);
480                                 record->resv = NULL;
481                         }
482                         record->data = NULL;
483                         --hammer_count_records;
484                         kfree(record, hmp->m_misc);
485                 }
486         }
487 }
488
489 /*
490  * Record visibility depends on whether the record is being accessed by
491  * the backend or the frontend.  Backend tests ignore the frontend delete
492  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
493  * must also check for commit races.
494  *
495  * Return non-zero if the record is visible, zero if it isn't or if it is
496  * deleted.  Returns 0 if the record has been comitted (unless the special
497  * delete-visibility flag is set).  A committed record must be located
498  * via the media B-Tree.  Returns non-zero if the record is good.
499  *
500  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
501  * records to be returned.  This is so pending deletions are detected
502  * when using an iterator to locate an unused hash key, or when we need
503  * to locate historical records on-disk to destroy.
504  */
505 static __inline
506 int
507 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
508 {
509         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
510                 return(1);
511         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
512                 if (record->flags & (HAMMER_RECF_DELETED_BE |
513                                      HAMMER_RECF_COMMITTED)) {
514                         return(0);
515                 }
516         } else {
517                 if (record->flags & (HAMMER_RECF_DELETED_FE |
518                                      HAMMER_RECF_DELETED_BE |
519                                      HAMMER_RECF_COMMITTED)) {
520                         return(0);
521                 }
522         }
523         return(1);
524 }
525
526 /*
527  * This callback is used as part of the RB_SCAN function for in-memory
528  * records.  We terminate it (return -1) as soon as we get a match.
529  *
530  * This routine is used by frontend code.
531  *
532  * The primary compare code does not account for ASOF lookups.  This
533  * code handles that case as well as a few others.
534  */
535 static
536 int
537 hammer_rec_scan_callback(hammer_record_t rec, void *data)
538 {
539         hammer_cursor_t cursor = data;
540
541         /*
542          * We terminate on success, so this should be NULL on entry.
543          */
544         KKASSERT(cursor->iprec == NULL);
545
546         /*
547          * Skip if the record was marked deleted or committed.
548          */
549         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
550                 return(0);
551
552         /*
553          * Skip if not visible due to our as-of TID
554          */
555         if (cursor->flags & HAMMER_CURSOR_ASOF) {
556                 if (cursor->asof < rec->leaf.base.create_tid)
557                         return(0);
558                 if (rec->leaf.base.delete_tid &&
559                     cursor->asof >= rec->leaf.base.delete_tid) {
560                         return(0);
561                 }
562         }
563
564         /*
565          * ref the record.  The record is protected from backend B-Tree
566          * interactions by virtue of the cursor's IP lock.
567          */
568         hammer_ref(&rec->lock);
569
570         /*
571          * The record may have been deleted or committed while we
572          * were blocked.  XXX remove?
573          */
574         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
575                 hammer_rel_mem_record(rec);
576                 return(0);
577         }
578
579         /*
580          * Set the matching record and stop the scan.
581          */
582         cursor->iprec = rec;
583         return(-1);
584 }
585
586
587 /*
588  * Lookup an in-memory record given the key specified in the cursor.  Works
589  * just like hammer_btree_lookup() but operates on an inode's in-memory
590  * record list.
591  *
592  * The lookup must fail if the record is marked for deferred deletion.
593  *
594  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
595  */
596 static
597 int
598 hammer_mem_lookup(hammer_cursor_t cursor)
599 {
600         KKASSERT(cursor->ip);
601         if (cursor->iprec) {
602                 hammer_rel_mem_record(cursor->iprec);
603                 cursor->iprec = NULL;
604         }
605         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
606                                    hammer_rec_scan_callback, cursor);
607
608         return (cursor->iprec ? 0 : ENOENT);
609 }
610
611 /*
612  * hammer_mem_first() - locate the first in-memory record matching the
613  * cursor within the bounds of the key range.
614  *
615  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
616  * will set ATEMEM the same as MEMEOF, and does not return any error.
617  */
618 static
619 void
620 hammer_mem_first(hammer_cursor_t cursor)
621 {
622         hammer_inode_t ip;
623
624         ip = cursor->ip;
625         KKASSERT(ip != NULL);
626
627         if (cursor->iprec) {
628                 hammer_rel_mem_record(cursor->iprec);
629                 cursor->iprec = NULL;
630         }
631         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
632                                    hammer_rec_scan_callback, cursor);
633
634         if (cursor->iprec)
635                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
636         else
637                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
638 }
639
640 /************************************************************************
641  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
642  ************************************************************************
643  *
644  * These functions manipulate in-memory records.  Such records typically
645  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
646  */
647
648 /*
649  * Add a directory entry (dip,ncp) which references inode (ip).
650  *
651  * Note that the low 32 bits of the namekey are set temporarily to create
652  * a unique in-memory record, and may be modified a second time when the
653  * record is synchronized to disk.  In particular, the low 32 bits cannot be
654  * all 0's when synching to disk, which is not handled here.
655  *
656  * NOTE: bytes does not include any terminating \0 on name, and name might
657  * not be terminated.
658  */
659 int
660 hammer_ip_add_directory(struct hammer_transaction *trans,
661                      struct hammer_inode *dip, const char *name, int bytes,
662                      struct hammer_inode *ip)
663 {
664         struct hammer_cursor cursor;
665         hammer_record_t record;
666         int error;
667         u_int32_t max_iterations;
668
669         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
670
671         record->type = HAMMER_MEM_RECORD_ADD;
672         record->leaf.base.localization = dip->obj_localization +
673                                          hammer_dir_localization(dip);
674         record->leaf.base.obj_id = dip->obj_id;
675         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
676                                                          &max_iterations);
677         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
678         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
679         record->data->entry.obj_id = ip->obj_id;
680         record->data->entry.localization = ip->obj_localization;
681         bcopy(name, record->data->entry.name, bytes);
682
683         ++ip->ino_data.nlinks;
684         ip->ino_data.ctime = trans->time;
685         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
686
687         /*
688          * Find an unused namekey.  Both the in-memory record tree and
689          * the B-Tree are checked.  We do not want historically deleted
690          * names to create a collision as our iteration space may be limited,
691          * and since create_tid wouldn't match anyway an ASOF search
692          * must be used to locate collisions.
693          *
694          * delete-visibility is set so pending deletions do not give us
695          * a false-negative on our ability to use an iterator.
696          *
697          * The iterator must not rollover the key.  Directory keys only
698          * use the positive key space.
699          */
700         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
701         cursor.key_beg = record->leaf.base;
702         cursor.flags |= HAMMER_CURSOR_ASOF;
703         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
704         cursor.asof = ip->obj_asof;
705
706         while (hammer_ip_lookup(&cursor) == 0) {
707                 ++record->leaf.base.key;
708                 KKASSERT(record->leaf.base.key > 0);
709                 cursor.key_beg.key = record->leaf.base.key;
710                 if (--max_iterations == 0) {
711                         hammer_rel_mem_record(record);
712                         error = ENOSPC;
713                         goto failed;
714                 }
715         }
716
717         /*
718          * The target inode and the directory entry are bound together.
719          */
720         record->target_ip = ip;
721         record->flush_state = HAMMER_FST_SETUP;
722         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
723
724         /*
725          * The inode now has a dependancy and must be taken out of the idle
726          * state.  An inode not in an idle state is given an extra reference.
727          *
728          * When transitioning to a SETUP state flag for an automatic reflush
729          * when the dependancies are disposed of if someone is waiting on
730          * the inode.
731          */
732         if (ip->flush_state == HAMMER_FST_IDLE) {
733                 hammer_ref(&ip->lock);
734                 ip->flush_state = HAMMER_FST_SETUP;
735                 if (ip->flags & HAMMER_INODE_FLUSHW)
736                         ip->flags |= HAMMER_INODE_REFLUSH;
737         }
738         error = hammer_mem_add(record);
739         if (error == 0) {
740                 dip->ino_data.mtime = trans->time;
741                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
742         }
743 failed:
744         hammer_done_cursor(&cursor);
745         return(error);
746 }
747
748 /*
749  * Delete the directory entry and update the inode link count.  The
750  * cursor must be seeked to the directory entry record being deleted.
751  *
752  * The related inode should be share-locked by the caller.  The caller is
753  * on the frontend.  It could also be NULL indicating that the directory
754  * entry being removed has no related inode.
755  *
756  * This function can return EDEADLK requiring the caller to terminate
757  * the cursor, any locks, wait on the returned record, and retry.
758  */
759 int
760 hammer_ip_del_directory(struct hammer_transaction *trans,
761                      hammer_cursor_t cursor, struct hammer_inode *dip,
762                      struct hammer_inode *ip)
763 {
764         hammer_record_t record;
765         int error;
766
767         if (hammer_cursor_inmem(cursor)) {
768                 /*
769                  * In-memory (unsynchronized) records can simply be freed.
770                  *
771                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
772                  * by the backend, we must still avoid races against the
773                  * backend potentially syncing the record to the media.
774                  *
775                  * We cannot call hammer_ip_delete_record(), that routine may
776                  * only be called from the backend.
777                  */
778                 record = cursor->iprec;
779                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
780                                      HAMMER_RECF_DELETED_BE |
781                                      HAMMER_RECF_COMMITTED)) {
782                         KKASSERT(cursor->deadlk_rec == NULL);
783                         hammer_ref(&record->lock);
784                         cursor->deadlk_rec = record;
785                         error = EDEADLK;
786                 } else {
787                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
788                         record->flags |= HAMMER_RECF_DELETED_FE;
789                         error = 0;
790                 }
791         } else {
792                 /*
793                  * If the record is on-disk we have to queue the deletion by
794                  * the record's key.  This also causes lookups to skip the
795                  * record (lookups for the purposes of finding an unused
796                  * directory key do not skip the record).
797                  */
798                 KKASSERT(dip->flags &
799                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
800                 record = hammer_alloc_mem_record(dip, 0);
801                 record->type = HAMMER_MEM_RECORD_DEL;
802                 record->leaf.base = cursor->leaf->base;
803                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
804
805                 /*
806                  * ip may be NULL, indicating the deletion of a directory
807                  * entry which has no related inode.
808                  */
809                 record->target_ip = ip;
810                 if (ip) {
811                         record->flush_state = HAMMER_FST_SETUP;
812                         TAILQ_INSERT_TAIL(&ip->target_list, record,
813                                           target_entry);
814                 } else {
815                         record->flush_state = HAMMER_FST_IDLE;
816                 }
817
818                 /*
819                  * The inode now has a dependancy and must be taken out of
820                  * the idle state.  An inode not in an idle state is given
821                  * an extra reference.
822                  *
823                  * When transitioning to a SETUP state flag for an automatic
824                  * reflush when the dependancies are disposed of if someone
825                  * is waiting on the inode.
826                  */
827                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
828                         hammer_ref(&ip->lock);
829                         ip->flush_state = HAMMER_FST_SETUP;
830                         if (ip->flags & HAMMER_INODE_FLUSHW)
831                                 ip->flags |= HAMMER_INODE_REFLUSH;
832                 }
833
834                 error = hammer_mem_add(record);
835         }
836
837         /*
838          * One less link.  The file may still be open in the OS even after
839          * all links have gone away.
840          *
841          * We have to terminate the cursor before syncing the inode to
842          * avoid deadlocking against ourselves.  XXX this may no longer
843          * be true.
844          *
845          * If nlinks drops to zero and the vnode is inactive (or there is
846          * no vnode), call hammer_inode_unloadable_check() to zonk the
847          * inode.  If we don't do this here the inode will not be destroyed
848          * on-media until we unmount.
849          */
850         if (error == 0) {
851                 if (ip) {
852                         --ip->ino_data.nlinks;  /* do before we might block */
853                         ip->ino_data.ctime = trans->time;
854                 }
855                 dip->ino_data.mtime = trans->time;
856                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
857                 if (ip) {
858                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
859                         if (ip->ino_data.nlinks == 0 &&
860                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
861                                 hammer_done_cursor(cursor);
862                                 hammer_inode_unloadable_check(ip, 1);
863                                 hammer_flush_inode(ip, 0);
864                         }
865                 }
866
867         }
868         return(error);
869 }
870
871 /*
872  * Add a record to an inode.
873  *
874  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
875  * initialize the following additional fields:
876  *
877  * The related inode should be share-locked by the caller.  The caller is
878  * on the frontend.
879  *
880  * record->rec.entry.base.base.key
881  * record->rec.entry.base.base.rec_type
882  * record->rec.entry.base.base.data_len
883  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
884  */
885 int
886 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
887 {
888         hammer_inode_t ip = record->ip;
889         int error;
890
891         KKASSERT(record->leaf.base.localization != 0);
892         record->leaf.base.obj_id = ip->obj_id;
893         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
894         error = hammer_mem_add(record);
895         return(error);
896 }
897
898 /*
899  * Locate a pre-existing bulk record in memory.  The caller wishes to
900  * replace the record with a new one.  The existing record may have a
901  * different length (and thus a different key) so we have to use an
902  * overlap check function.
903  */
904 static hammer_record_t
905 hammer_ip_get_bulk(hammer_record_t record)
906 {
907         struct hammer_bulk_info info;
908         hammer_inode_t ip = record->ip;
909
910         info.record = record;
911         info.conflict = NULL;
912         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
913                                    hammer_bulk_scan_callback, &info);
914
915         return(info.conflict);  /* may be NULL */
916 }
917
918 /*
919  * Take records vetted by overlap_cmp.  The first non-deleted record
920  * (if any) stops the scan.
921  */
922 static int
923 hammer_bulk_scan_callback(hammer_record_t record, void *data)
924 {
925         struct hammer_bulk_info *info = data;
926
927         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
928                              HAMMER_RECF_COMMITTED)) {
929                 return(0);
930         }
931         hammer_ref(&record->lock);
932         info->conflict = record;
933         return(-1);                     /* stop scan */
934 }
935
936 /*
937  * Reserve blockmap space placemarked with an in-memory record.  
938  *
939  * This routine is called by the frontend in order to be able to directly
940  * flush a buffer cache buffer.  The frontend has locked the related buffer
941  * cache buffers and we should be able to manipulate any overlapping
942  * in-memory records.
943  *
944  * The caller is responsible for adding the returned record and deleting
945  * the returned conflicting record (if any), typically by calling
946  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
947  */
948 hammer_record_t
949 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
950                    int *errorp)
951 {
952         hammer_record_t record;
953         hammer_dedup_cache_t dcp;
954         hammer_crc_t crc;
955         int zone;
956
957         /*
958          * Create a record to cover the direct write.  The record cannot
959          * be added to the in-memory RB tree here as it might conflict
960          * with an existing memory record.  See hammer_io_direct_write().
961          *
962          * The backend is responsible for finalizing the space reserved in
963          * this record.
964          *
965          * XXX bytes not aligned, depend on the reservation code to
966          * align the reservation.
967          */
968         record = hammer_alloc_mem_record(ip, 0);
969         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
970                                            HAMMER_ZONE_SMALL_DATA_INDEX;
971         if (bytes == 0)
972                 crc = 0;
973         else
974                 crc = crc32(data, bytes);
975
976         if (hammer_live_dedup == 0)
977                 goto nodedup;
978         if ((dcp = hammer_dedup_cache_lookup(ip->hmp, crc)) != NULL) {
979                 struct hammer_dedup_cache tmp = *dcp;
980
981                 record->resv = hammer_blockmap_reserve_dedup(ip->hmp, zone,
982                         bytes, tmp.data_offset, errorp);
983                 if (record->resv == NULL)
984                         goto nodedup;
985
986                 if (!hammer_dedup_validate(&tmp, zone, bytes, data)) {
987                         hammer_blockmap_reserve_complete(ip->hmp, record->resv);
988                         goto nodedup;
989                 }
990
991                 record->leaf.data_offset = tmp.data_offset;
992                 record->flags |= HAMMER_RECF_DEDUPED;
993         } else {
994 nodedup:
995                 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
996                        &record->leaf.data_offset, errorp);
997                 if (record->resv == NULL) {
998                         kprintf("hammer_ip_add_bulk: reservation failed\n");
999                         hammer_rel_mem_record(record);
1000                         return(NULL);
1001                 }
1002         }
1003
1004         record->type = HAMMER_MEM_RECORD_DATA;
1005         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
1006         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
1007         record->leaf.base.obj_id = ip->obj_id;
1008         record->leaf.base.key = file_offset + bytes;
1009         record->leaf.base.localization = ip->obj_localization +
1010                                          HAMMER_LOCALIZE_MISC;
1011         record->leaf.data_len = bytes;
1012         record->leaf.data_crc = crc;
1013         KKASSERT(*errorp == 0);
1014
1015         return(record);
1016 }
1017
1018 /*
1019  * Called by hammer_io_direct_write() prior to any possible completion
1020  * of the BIO to emplace the memory record associated with the I/O and
1021  * to replace any prior memory record which might still be active.
1022  *
1023  * Setting the FE deleted flag on the old record (if any) avoids any RB
1024  * tree insertion conflict, amoung other things.
1025  *
1026  * This has to be done prior to the caller completing any related buffer
1027  * cache I/O or a reinstantiation of the buffer may load data from the
1028  * old media location instead of the new media location.  The holding
1029  * of the locked buffer cache buffer serves to interlock the record
1030  * replacement operation.
1031  */
1032 void
1033 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1034 {
1035         hammer_record_t conflict;
1036         int error;
1037
1038         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1039                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1040                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1041                         break;
1042                 }
1043                 conflict->flags |= HAMMER_RECF_WANTED;
1044                 tsleep(conflict, 0, "hmrrc3", 0);
1045                 hammer_rel_mem_record(conflict);
1046         }
1047         error = hammer_mem_add(record);
1048         if (conflict)
1049                 hammer_rel_mem_record(conflict);
1050         KKASSERT(error == 0);
1051 }
1052
1053 /*
1054  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1055  * and records in a flushing state are handled by the backend.  The vnops
1056  * setattr code will handle the block containing the truncation point.
1057  *
1058  * Partial blocks are not deleted.
1059  *
1060  * This code is only called on regular files.
1061  */
1062 int
1063 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1064 {
1065         struct rec_trunc_info info;
1066
1067         switch(ip->ino_data.obj_type) {
1068         case HAMMER_OBJTYPE_REGFILE:
1069                 info.rec_type = HAMMER_RECTYPE_DATA;
1070                 break;
1071         case HAMMER_OBJTYPE_DBFILE:
1072                 info.rec_type = HAMMER_RECTYPE_DB;
1073                 break;
1074         default:
1075                 return(EINVAL);
1076         }
1077         info.trunc_off = file_size;
1078         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1079                                    hammer_frontend_trunc_callback, &info);
1080         return(0);
1081 }
1082
1083 /*
1084  * Scan callback for frontend records to destroy during a truncation.
1085  * We must ensure that DELETED_FE is set on the record or the frontend
1086  * will get confused in future read() calls.
1087  *
1088  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1089  *       In this rare case we must wait for the interlock to be cleared.
1090  *
1091  * NOTE: This function is only called on regular files.  There are further
1092  *       restrictions to the setting of DELETED_FE on directory records
1093  *       undergoing a flush due to sensitive inode link count calculations.
1094  */
1095 static int
1096 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1097 {
1098         if (record->flags & HAMMER_RECF_DELETED_FE)
1099                 return(0);
1100 #if 0
1101         if (record->flush_state == HAMMER_FST_FLUSH)
1102                 return(0);
1103 #endif
1104         hammer_ref(&record->lock);
1105         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1106                 hammer_wait_mem_record_ident(record, "hmmtrr");
1107         record->flags |= HAMMER_RECF_DELETED_FE;
1108         hammer_rel_mem_record(record);
1109         return(0);
1110 }
1111
1112 /*
1113  * Return 1 if the caller must check for and delete existing records
1114  * before writing out a new data record.
1115  *
1116  * Return 0 if the caller can just insert the record into the B-Tree without
1117  * checking.
1118  */
1119 static int
1120 hammer_record_needs_overwrite_delete(hammer_record_t record)
1121 {
1122         hammer_inode_t ip = record->ip;
1123         int64_t file_offset;
1124         int r;
1125
1126         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1127                 file_offset = record->leaf.base.key;
1128         else
1129                 file_offset = record->leaf.base.key - record->leaf.data_len;
1130         r = (file_offset < ip->save_trunc_off);
1131         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1132                 if (ip->save_trunc_off <= record->leaf.base.key)
1133                         ip->save_trunc_off = record->leaf.base.key + 1;
1134         } else {
1135                 if (ip->save_trunc_off < record->leaf.base.key)
1136                         ip->save_trunc_off = record->leaf.base.key;
1137         }
1138         return(r);
1139 }
1140
1141 /*
1142  * Backend code.  Sync a record to the media.
1143  */
1144 int
1145 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1146 {
1147         hammer_transaction_t trans = cursor->trans;
1148         int64_t file_offset;
1149         int bytes;
1150         void *bdata;
1151         int error;
1152         int doprop;
1153
1154         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1155         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1156         KKASSERT(record->leaf.base.localization != 0);
1157
1158         /*
1159          * Any direct-write related to the record must complete before we
1160          * can sync the record to the on-disk media.
1161          */
1162         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1163                 hammer_io_direct_wait(record);
1164
1165         /*
1166          * If this is a bulk-data record placemarker there may be an existing
1167          * record on-disk, indicating a data overwrite.  If there is the
1168          * on-disk record must be deleted before we can insert our new record.
1169          *
1170          * We've synthesized this record and do not know what the create_tid
1171          * on-disk is, nor how much data it represents.
1172          *
1173          * Keep in mind that (key) for data records is (base_offset + len),
1174          * not (base_offset).  Also, we only want to get rid of on-disk
1175          * records since we are trying to sync our in-memory record, call
1176          * hammer_ip_delete_range() with truncating set to 1 to make sure
1177          * it skips in-memory records.
1178          *
1179          * It is ok for the lookup to return ENOENT.
1180          *
1181          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1182          * to call hammer_ip_delete_range() or not.  This also means we must
1183          * update sync_trunc_off() as we write.
1184          */
1185         if (record->type == HAMMER_MEM_RECORD_DATA &&
1186             hammer_record_needs_overwrite_delete(record)) {
1187                 file_offset = record->leaf.base.key - record->leaf.data_len;
1188                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) & 
1189                         ~HAMMER_BUFMASK;
1190                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1191                 error = hammer_ip_delete_range(
1192                                 cursor, record->ip,
1193                                 file_offset, file_offset + bytes - 1,
1194                                 1);
1195                 if (error && error != ENOENT)
1196                         goto done;
1197         }
1198
1199         /*
1200          * If this is a general record there may be an on-disk version
1201          * that must be deleted before we can insert the new record.
1202          */
1203         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1204                 error = hammer_delete_general(cursor, record->ip,
1205                                               &record->leaf);
1206                 if (error && error != ENOENT)
1207                         goto done;
1208         }
1209
1210         /*
1211          * Setup the cursor.
1212          */
1213         hammer_normalize_cursor(cursor);
1214         cursor->key_beg = record->leaf.base;
1215         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1216         cursor->flags |= HAMMER_CURSOR_BACKEND;
1217         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1218
1219         /*
1220          * Records can wind up on-media before the inode itself is on-media.
1221          * Flag the case.
1222          */
1223         record->ip->flags |= HAMMER_INODE_DONDISK;
1224
1225         /*
1226          * If we are deleting a directory entry an exact match must be
1227          * found on-disk.
1228          */
1229         if (record->type == HAMMER_MEM_RECORD_DEL) {
1230                 error = hammer_btree_lookup(cursor);
1231                 if (error == 0) {
1232                         KKASSERT(cursor->iprec == NULL);
1233                         error = hammer_ip_delete_record(cursor, record->ip,
1234                                                         trans->tid);
1235                         if (error == 0) {
1236                                 record->flags |= HAMMER_RECF_DELETED_BE |
1237                                                  HAMMER_RECF_COMMITTED;
1238                                 ++record->ip->rec_generation;
1239                         }
1240                 }
1241                 goto done;
1242         }
1243
1244         /*
1245          * We are inserting.
1246          *
1247          * Issue a lookup to position the cursor and locate the insertion
1248          * point.  The target key should not exist.  If we are creating a
1249          * directory entry we may have to iterate the low 32 bits of the
1250          * key to find an unused key.
1251          */
1252         hammer_sync_lock_sh(trans);
1253         cursor->flags |= HAMMER_CURSOR_INSERT;
1254         error = hammer_btree_lookup(cursor);
1255         if (hammer_debug_inode)
1256                 kprintf("DOINSERT LOOKUP %d\n", error);
1257         if (error == 0) {
1258                 kprintf("hammer_ip_sync_record: duplicate rec "
1259                         "at (%016llx)\n", (long long)record->leaf.base.key);
1260                 if (hammer_debug_critical)
1261                         Debugger("duplicate record1");
1262                 error = EIO;
1263         }
1264 #if 0
1265         if (record->type == HAMMER_MEM_RECORD_DATA)
1266                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1267                         record->leaf.base.key - record->leaf.data_len,
1268                         record->leaf.data_offset, error);
1269 #endif
1270
1271         if (error != ENOENT)
1272                 goto done_unlock;
1273
1274         /*
1275          * Allocate the record and data.  The result buffers will be
1276          * marked as being modified and further calls to
1277          * hammer_modify_buffer() will result in unneeded UNDO records.
1278          *
1279          * Support zero-fill records (data == NULL and data_len != 0)
1280          */
1281         if (record->type == HAMMER_MEM_RECORD_DATA) {
1282                 /*
1283                  * The data portion of a bulk-data record has already been
1284                  * committed to disk, we need only adjust the layer2
1285                  * statistics in the same transaction as our B-Tree insert.
1286                  */
1287                 KKASSERT(record->leaf.data_offset != 0);
1288                 error = hammer_blockmap_finalize(trans,
1289                                                  record->resv,
1290                                                  record->leaf.data_offset,
1291                                                  record->leaf.data_len);
1292
1293                 if (hammer_live_dedup == 2 &&
1294                     (record->flags & HAMMER_RECF_DEDUPED) == 0) {
1295                         hammer_dedup_cache_add(record->ip, &record->leaf);
1296                 }
1297         } else if (record->data && record->leaf.data_len) {
1298                 /*
1299                  * Wholely cached record, with data.  Allocate the data.
1300                  */
1301                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1302                                           record->leaf.base.rec_type,
1303                                           &record->leaf.data_offset,
1304                                           &cursor->data_buffer,
1305                                           0, &error);
1306                 if (bdata == NULL)
1307                         goto done_unlock;
1308                 hammer_crc_set_leaf(record->data, &record->leaf);
1309                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1310                 bcopy(record->data, bdata, record->leaf.data_len);
1311                 hammer_modify_buffer_done(cursor->data_buffer);
1312         } else {
1313                 /*
1314                  * Wholely cached record, without data.
1315                  */
1316                 record->leaf.data_offset = 0;
1317                 record->leaf.data_crc = 0;
1318         }
1319
1320         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1321         if (hammer_debug_inode && error) {
1322                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1323                         error,
1324                         (long long)cursor->node->node_offset,
1325                         cursor->index,
1326                         (long long)record->leaf.base.key);
1327         }
1328
1329         /*
1330          * Our record is on-disk and we normally mark the in-memory version
1331          * as having been committed (and not BE-deleted).
1332          *
1333          * If the record represented a directory deletion but we had to
1334          * sync a valid directory entry to disk due to dependancies,
1335          * we must convert the record to a covering delete so the
1336          * frontend does not have visibility on the synced entry.
1337          *
1338          * WARNING: cursor's leaf pointer may have changed after do_propagation
1339          *          returns!
1340          */
1341         if (error == 0) {
1342                 if (doprop) {
1343                         hammer_btree_do_propagation(cursor,
1344                                                     record->ip->pfsm,
1345                                                     &record->leaf);
1346                 }
1347                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1348                         /*
1349                          * Must convert deleted directory entry add
1350                          * to a directory entry delete.
1351                          */
1352                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1353                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1354                         record->type = HAMMER_MEM_RECORD_DEL;
1355                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1356                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1357                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1358                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1359                                                  HAMMER_RECF_DELETED_BE)) == 0);
1360                         /* converted record is not yet committed */
1361                         /* hammer_flush_record_done takes care of the rest */
1362                 } else {
1363                         /*
1364                          * Everything went fine and we are now done with
1365                          * this record.
1366                          */
1367                         record->flags |= HAMMER_RECF_COMMITTED;
1368                         ++record->ip->rec_generation;
1369                 }
1370         } else {
1371                 if (record->leaf.data_offset) {
1372                         hammer_blockmap_free(trans, record->leaf.data_offset,
1373                                              record->leaf.data_len);
1374                 }
1375         }
1376 done_unlock:
1377         hammer_sync_unlock(trans);
1378 done:
1379         return(error);
1380 }
1381
1382 /*
1383  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1384  * entry's key is used to deal with hash collisions in the upper 32 bits.
1385  * A unique 64 bit key is generated in-memory and may be regenerated a
1386  * second time when the directory record is flushed to the on-disk B-Tree.
1387  *
1388  * A referenced record is passed to this function.  This function
1389  * eats the reference.  If an error occurs the record will be deleted.
1390  *
1391  * A copy of the temporary record->data pointer provided by the caller
1392  * will be made.
1393  */
1394 int
1395 hammer_mem_add(hammer_record_t record)
1396 {
1397         hammer_mount_t hmp = record->ip->hmp;
1398
1399         /*
1400          * Make a private copy of record->data
1401          */
1402         if (record->data)
1403                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1404
1405         /*
1406          * Insert into the RB tree.  A unique key should have already
1407          * been selected if this is a directory entry.
1408          */
1409         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1410                 record->flags |= HAMMER_RECF_DELETED_FE;
1411                 hammer_rel_mem_record(record);
1412                 return (EEXIST);
1413         }
1414         ++hmp->count_newrecords;
1415         ++hmp->rsv_recs;
1416         ++record->ip->rsv_recs;
1417         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1418         record->flags |= HAMMER_RECF_ONRBTREE;
1419         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1420         hammer_rel_mem_record(record);
1421         return(0);
1422 }
1423
1424 /************************************************************************
1425  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1426  ************************************************************************
1427  *
1428  * These functions augment the B-Tree scanning functions in hammer_btree.c
1429  * by merging in-memory records with on-disk records.
1430  */
1431
1432 /*
1433  * Locate a particular record either in-memory or on-disk.
1434  *
1435  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1436  * NOT be called to iterate results.
1437  */
1438 int
1439 hammer_ip_lookup(hammer_cursor_t cursor)
1440 {
1441         int error;
1442
1443         /*
1444          * If the element is in-memory return it without searching the
1445          * on-disk B-Tree
1446          */
1447         KKASSERT(cursor->ip);
1448         error = hammer_mem_lookup(cursor);
1449         if (error == 0) {
1450                 cursor->leaf = &cursor->iprec->leaf;
1451                 return(error);
1452         }
1453         if (error != ENOENT)
1454                 return(error);
1455
1456         /*
1457          * If the inode has on-disk components search the on-disk B-Tree.
1458          */
1459         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1460                 return(error);
1461         error = hammer_btree_lookup(cursor);
1462         if (error == 0)
1463                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1464         return(error);
1465 }
1466
1467 /*
1468  * Helper for hammer_ip_first()/hammer_ip_next()
1469  *
1470  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1471  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1472  * state if hammer_ip_next() needs to re-seek.
1473  */
1474 static __inline
1475 int
1476 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1477 {
1478         hammer_inode_t ip = cursor->ip;
1479         int error;
1480
1481         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1482                 error = hammer_btree_lookup(cursor);
1483                 if (error == ENOENT || error == EDEADLK) {
1484                         if (hammer_debug_general & 0x2000) {
1485                                 kprintf("error %d node %p %016llx index %d\n",
1486                                         error, cursor->node,
1487                                         (long long)cursor->node->node_offset,
1488                                         cursor->index);
1489                         }
1490                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1491                         error = hammer_btree_iterate(cursor);
1492                 }
1493                 if (error == 0) {
1494                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1495                                            HAMMER_CURSOR_ATEDISK);
1496                 } else {
1497                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1498                                          HAMMER_CURSOR_ATEDISK;
1499                         if (error == ENOENT)
1500                                 error = 0;
1501                 }
1502         } else {
1503                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1504                 error = 0;
1505         }
1506         return(error);
1507 }
1508
1509 /*
1510  * Helper for hammer_ip_next()
1511  *
1512  * The caller has determined that the media cursor is further along than the
1513  * memory cursor and must be reseeked after a generation number change.
1514  */
1515 static
1516 int
1517 _hammer_ip_reseek(hammer_cursor_t cursor)
1518 {
1519         struct hammer_base_elm save;
1520         hammer_btree_elm_t elm;
1521         int error;
1522         int r;
1523         int again = 0;
1524
1525         /*
1526          * Do the re-seek.
1527          */
1528         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1529                 (long long)cursor->ip->obj_id);
1530         save = cursor->key_beg;
1531         cursor->key_beg = cursor->iprec->leaf.base;
1532         error = _hammer_ip_seek_btree(cursor);
1533         KKASSERT(error == 0);
1534         cursor->key_beg = save;
1535
1536         /*
1537          * If the memory record was previous returned to
1538          * the caller and the media record matches
1539          * (-1/+1: only create_tid differs), then iterate
1540          * the media record to avoid a double result.
1541          */
1542         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1543             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1544                 elm = &cursor->node->ondisk->elms[cursor->index];
1545                 r = hammer_btree_cmp(&elm->base,
1546                                      &cursor->iprec->leaf.base);
1547                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1548                         if (r >= -1 && r <= 1) {
1549                                 kprintf("HAMMER: Debug: iterated after "
1550                                         "re-seek (asof r=%d)\n", r);
1551                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1552                                 again = 1;
1553                         }
1554                 } else {
1555                         if (r == 0) {
1556                                 kprintf("HAMMER: Debug: iterated after "
1557                                         "re-seek\n");
1558                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1559                                 again = 1;
1560                         }
1561                 }
1562         }
1563         return(again);
1564 }
1565
1566 /*
1567  * Locate the first record within the cursor's key_beg/key_end range,
1568  * restricted to a particular inode.  0 is returned on success, ENOENT
1569  * if no records matched the requested range, or some other error.
1570  *
1571  * When 0 is returned hammer_ip_next() may be used to iterate additional
1572  * records within the requested range.
1573  *
1574  * This function can return EDEADLK, requiring the caller to terminate
1575  * the cursor and try again.
1576  */
1577
1578 int
1579 hammer_ip_first(hammer_cursor_t cursor)
1580 {
1581         hammer_inode_t ip __debugvar = cursor->ip;
1582         int error;
1583
1584         KKASSERT(ip != NULL);
1585
1586         /*
1587          * Clean up fields and setup for merged scan
1588          */
1589         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1590
1591         /*
1592          * Search the in-memory record list (Red-Black tree).  Unlike the
1593          * B-Tree search, mem_first checks for records in the range.
1594          *
1595          * This function will setup both ATEMEM and MEMEOF properly for
1596          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1597          */
1598         hammer_mem_first(cursor);
1599
1600         /*
1601          * Detect generation changes during blockages, including
1602          * blockages which occur on the initial btree search.
1603          */
1604         cursor->rec_generation = cursor->ip->rec_generation;
1605
1606         /*
1607          * Initial search and result
1608          */
1609         error = _hammer_ip_seek_btree(cursor);
1610         if (error == 0)
1611                 error = hammer_ip_next(cursor);
1612
1613         return (error);
1614 }
1615
1616 /*
1617  * Retrieve the next record in a merged iteration within the bounds of the
1618  * cursor.  This call may be made multiple times after the cursor has been
1619  * initially searched with hammer_ip_first().
1620  *
1621  * There are numerous special cases in this code to deal with races between
1622  * in-memory records and on-media records.
1623  *
1624  * 0 is returned on success, ENOENT if no further records match the
1625  * requested range, or some other error code is returned.
1626  */
1627 int
1628 hammer_ip_next(hammer_cursor_t cursor)
1629 {
1630         hammer_btree_elm_t elm;
1631         hammer_record_t rec;
1632         hammer_record_t tmprec;
1633         int error;
1634         int r;
1635
1636 again:
1637         /*
1638          * Get the next on-disk record
1639          *
1640          * NOTE: If we deleted the last on-disk record we had scanned
1641          *       ATEDISK will be clear and RETEST will be set, forcing
1642          *       a call to iterate.  The fact that ATEDISK is clear causes
1643          *       iterate to re-test the 'current' element.  If ATEDISK is
1644          *       set, iterate will skip the 'current' element.
1645          */
1646         error = 0;
1647         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1648                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1649                                      HAMMER_CURSOR_RETEST)) {
1650                         error = hammer_btree_iterate(cursor);
1651                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1652                         if (error == 0) {
1653                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1654                                 hammer_cache_node(&cursor->ip->cache[1],
1655                                                   cursor->node);
1656                         } else if (error == ENOENT) {
1657                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1658                                                  HAMMER_CURSOR_ATEDISK;
1659                                 error = 0;
1660                         }
1661                 }
1662         }
1663
1664         /*
1665          * If the generation changed the backend has deleted or committed
1666          * one or more memory records since our last check.
1667          *
1668          * When this case occurs if the disk cursor is > current memory record
1669          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1670          * Since the cursor is ahead it must have not yet been eaten (if
1671          * not at eof anyway). (XXX data offset case?)
1672          *
1673          * NOTE: we are not doing a full check here.  That will be handled
1674          * later on.
1675          *
1676          * If we have exhausted all memory records we do not have to do any
1677          * further seeks.
1678          */
1679         while (cursor->rec_generation != cursor->ip->rec_generation &&
1680                error == 0
1681         ) {
1682                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1683                 cursor->rec_generation = cursor->ip->rec_generation;
1684                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1685                         break;
1686                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1687                         r = 1;
1688                 } else {
1689                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1690                         elm = &cursor->node->ondisk->elms[cursor->index];
1691                         r = hammer_btree_cmp(&elm->base,
1692                                              &cursor->iprec->leaf.base);
1693                 }
1694
1695                 /*
1696                  * Do we re-seek the media cursor?
1697                  */
1698                 if (r > 0) {
1699                         if (_hammer_ip_reseek(cursor))
1700                                 goto again;
1701                 }
1702         }
1703
1704         /*
1705          * We can now safely get the next in-memory record.  We cannot
1706          * block here.
1707          *
1708          * hammer_rec_scan_cmp:  Is the record still in our general range,
1709          *                       (non-inclusive of snapshot exclusions)?
1710          * hammer_rec_scan_callback: Is the record in our snapshot?
1711          */
1712         tmprec = NULL;
1713         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1714                 /*
1715                  * If the current memory record was eaten then get the next
1716                  * one.  Stale records are skipped.
1717                  */
1718                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1719                         tmprec = cursor->iprec;
1720                         cursor->iprec = NULL;
1721                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1722                         while (rec) {
1723                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1724                                         break;
1725                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1726                                         break;
1727                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1728                         }
1729                         if (cursor->iprec) {
1730                                 KKASSERT(cursor->iprec == rec);
1731                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1732                         } else {
1733                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1734                         }
1735                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1736                 }
1737         }
1738
1739         /*
1740          * MEMORY RECORD VALIDITY TEST
1741          *
1742          * (We still can't block, which is why tmprec is being held so
1743          * long).
1744          *
1745          * If the memory record is no longer valid we skip it.  It may
1746          * have been deleted by the frontend.  If it was deleted or
1747          * committed by the backend the generation change re-seeked the
1748          * disk cursor and the record will be present there.
1749          */
1750         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1751                 KKASSERT(cursor->iprec);
1752                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1753                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1754                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1755                         if (tmprec)
1756                                 hammer_rel_mem_record(tmprec);
1757                         goto again;
1758                 }
1759         }
1760         if (tmprec)
1761                 hammer_rel_mem_record(tmprec);
1762
1763         /*
1764          * Extract either the disk or memory record depending on their
1765          * relative position.
1766          */
1767         error = 0;
1768         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1769         case 0:
1770                 /*
1771                  * Both entries valid.   Compare the entries and nominally
1772                  * return the first one in the sort order.  Numerous cases
1773                  * require special attention, however.
1774                  */
1775                 elm = &cursor->node->ondisk->elms[cursor->index];
1776                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1777
1778                 /*
1779                  * If the two entries differ only by their key (-2/2) or
1780                  * create_tid (-1/1), and are DATA records, we may have a
1781                  * nominal match.  We have to calculate the base file
1782                  * offset of the data.
1783                  */
1784                 if (r <= 2 && r >= -2 && r != 0 &&
1785                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1786                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1787                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1788                         int64_t base2 = cursor->iprec->leaf.base.key -
1789                                         cursor->iprec->leaf.data_len;
1790                         if (base1 == base2)
1791                                 r = 0;
1792                 }
1793
1794                 if (r < 0) {
1795                         error = hammer_btree_extract(cursor,
1796                                                      HAMMER_CURSOR_GET_LEAF);
1797                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1798                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1799                         break;
1800                 }
1801
1802                 /*
1803                  * If the entries match exactly the memory entry is either
1804                  * an on-disk directory entry deletion or a bulk data
1805                  * overwrite.  If it is a directory entry deletion we eat
1806                  * both entries.
1807                  *
1808                  * For the bulk-data overwrite case it is possible to have
1809                  * visibility into both, which simply means the syncer
1810                  * hasn't gotten around to doing the delete+insert sequence
1811                  * on the B-Tree.  Use the memory entry and throw away the
1812                  * on-disk entry.
1813                  *
1814                  * If the in-memory record is not either of these we
1815                  * probably caught the syncer while it was syncing it to
1816                  * the media.  Since we hold a shared lock on the cursor,
1817                  * the in-memory record had better be marked deleted at
1818                  * this point.
1819                  */
1820                 if (r == 0) {
1821                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1822                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1823                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1824                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1825                                         goto again;
1826                                 }
1827                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1828                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1829                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1830                                 }
1831                                 /* fall through to memory entry */
1832                         } else {
1833                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1834                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1835                                 goto again;
1836                         }
1837                 }
1838                 /* fall through to the memory entry */
1839         case HAMMER_CURSOR_ATEDISK:
1840                 /*
1841                  * Only the memory entry is valid.
1842                  */
1843                 cursor->leaf = &cursor->iprec->leaf;
1844                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1845                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1846
1847                 /*
1848                  * If the memory entry is an on-disk deletion we should have
1849                  * also had found a B-Tree record.  If the backend beat us
1850                  * to it it would have interlocked the cursor and we should
1851                  * have seen the in-memory record marked DELETED_FE.
1852                  */
1853                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1854                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1855                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1856                 }
1857                 break;
1858         case HAMMER_CURSOR_ATEMEM:
1859                 /*
1860                  * Only the disk entry is valid
1861                  */
1862                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1863                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1864                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1865                 break;
1866         default:
1867                 /*
1868                  * Neither entry is valid
1869                  *
1870                  * XXX error not set properly
1871                  */
1872                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1873                 cursor->leaf = NULL;
1874                 error = ENOENT;
1875                 break;
1876         }
1877         return(error);
1878 }
1879
1880 /*
1881  * Resolve the cursor->data pointer for the current cursor position in
1882  * a merged iteration.
1883  */
1884 int
1885 hammer_ip_resolve_data(hammer_cursor_t cursor)
1886 {
1887         hammer_record_t record;
1888         int error;
1889
1890         if (hammer_cursor_inmem(cursor)) {
1891                 /*
1892                  * The data associated with an in-memory record is usually
1893                  * kmalloced, but reserve-ahead data records will have an
1894                  * on-disk reference.
1895                  *
1896                  * NOTE: Reserve-ahead data records must be handled in the
1897                  * context of the related high level buffer cache buffer
1898                  * to interlock against async writes.
1899                  */
1900                 record = cursor->iprec;
1901                 cursor->data = record->data;
1902                 error = 0;
1903                 if (cursor->data == NULL) {
1904                         KKASSERT(record->leaf.base.rec_type ==
1905                                  HAMMER_RECTYPE_DATA);
1906                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1907                                                     record->leaf.data_offset,
1908                                                     record->leaf.data_len,
1909                                                     &error,
1910                                                     &cursor->data_buffer);
1911                 }
1912         } else {
1913                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1914                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1915         }
1916         return(error);
1917 }
1918
1919 /*
1920  * Backend truncation / record replacement - delete records in range.
1921  *
1922  * Delete all records within the specified range for inode ip.  In-memory
1923  * records still associated with the frontend are ignored. 
1924  *
1925  * If truncating is non-zero in-memory records associated with the back-end
1926  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1927  *
1928  * NOTES:
1929  *
1930  *      * An unaligned range will cause new records to be added to cover
1931  *        the edge cases. (XXX not implemented yet).
1932  *
1933  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1934  *        also do not deal with unaligned ranges.
1935  *
1936  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1937  *
1938  *      * Record keys for regular file data have to be special-cased since
1939  *        they indicate the end of the range (key = base + bytes).
1940  *
1941  *      * This function may be asked to delete ridiculously huge ranges, for
1942  *        example if someone truncates or removes a 1TB regular file.  We
1943  *        must be very careful on restarts and we may have to stop w/
1944  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1945  */
1946 int
1947 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1948                        int64_t ran_beg, int64_t ran_end, int truncating)
1949 {
1950         hammer_transaction_t trans = cursor->trans;
1951         hammer_btree_leaf_elm_t leaf;
1952         int error;
1953         int64_t off;
1954         int64_t tmp64;
1955
1956 #if 0
1957         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1958 #endif
1959
1960         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1961 retry:
1962         hammer_normalize_cursor(cursor);
1963         cursor->key_beg.localization = ip->obj_localization +
1964                                        HAMMER_LOCALIZE_MISC;
1965         cursor->key_beg.obj_id = ip->obj_id;
1966         cursor->key_beg.create_tid = 0;
1967         cursor->key_beg.delete_tid = 0;
1968         cursor->key_beg.obj_type = 0;
1969
1970         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1971                 cursor->key_beg.key = ran_beg;
1972                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1973         } else {
1974                 /*
1975                  * The key in the B-Tree is (base+bytes), so the first possible
1976                  * matching key is ran_beg + 1.
1977                  */
1978                 cursor->key_beg.key = ran_beg + 1;
1979                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1980         }
1981
1982         cursor->key_end = cursor->key_beg;
1983         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1984                 cursor->key_end.key = ran_end;
1985         } else {
1986                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1987                 if (tmp64 < ran_end)
1988                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1989                 else
1990                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1991         }
1992
1993         cursor->asof = ip->obj_asof;
1994         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1995         cursor->flags |= HAMMER_CURSOR_ASOF;
1996         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1997         cursor->flags |= HAMMER_CURSOR_BACKEND;
1998         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1999
2000         error = hammer_ip_first(cursor);
2001
2002         /*
2003          * Iterate through matching records and mark them as deleted.
2004          */
2005         while (error == 0) {
2006                 leaf = cursor->leaf;
2007
2008                 KKASSERT(leaf->base.delete_tid == 0);
2009                 KKASSERT(leaf->base.obj_id == ip->obj_id);
2010
2011                 /*
2012                  * There may be overlap cases for regular file data.  Also
2013                  * remember the key for a regular file record is (base + len),
2014                  * NOT (base).
2015                  *
2016                  * Note that due to duplicates (mem & media) allowed by
2017                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
2018                  */
2019                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
2020                         off = leaf->base.key - leaf->data_len;
2021                         /*
2022                          * Check the left edge case.  We currently do not
2023                          * split existing records.
2024                          */
2025                         if (off < ran_beg && leaf->base.key > ran_beg) {
2026                                 panic("hammer left edge case %016llx %d\n",
2027                                         (long long)leaf->base.key,
2028                                         leaf->data_len);
2029                         }
2030
2031                         /*
2032                          * Check the right edge case.  Note that the
2033                          * record can be completely out of bounds, which
2034                          * terminates the search.
2035                          *
2036                          * base->key is exclusive of the right edge while
2037                          * ran_end is inclusive of the right edge.  The
2038                          * (key - data_len) left boundary is inclusive.
2039                          *
2040                          * XXX theory-check this test at some point, are
2041                          * we missing a + 1 somewhere?  Note that ran_end
2042                          * could overflow.
2043                          */
2044                         if (leaf->base.key - 1 > ran_end) {
2045                                 if (leaf->base.key - leaf->data_len > ran_end)
2046                                         break;
2047                                 panic("hammer right edge case\n");
2048                         }
2049                 } else {
2050                         off = leaf->base.key;
2051                 }
2052
2053                 /*
2054                  * Delete the record.  When truncating we do not delete
2055                  * in-memory (data) records because they represent data
2056                  * written after the truncation.
2057                  *
2058                  * This will also physically destroy the B-Tree entry and
2059                  * data if the retention policy dictates.  The function
2060                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2061                  * to retest the new 'current' element.
2062                  */
2063                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2064                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2065                         /*
2066                          * If we have built up too many meta-buffers we risk
2067                          * deadlocking the kernel and must stop.  This can
2068                          * occur when deleting ridiculously huge files.
2069                          * sync_trunc_off is updated so the next cycle does
2070                          * not re-iterate records we have already deleted.
2071                          *
2072                          * This is only done with formal truncations.
2073                          */
2074                         if (truncating > 1 && error == 0 &&
2075                             hammer_flusher_meta_limit(ip->hmp)) {
2076                                 ip->sync_trunc_off = off;
2077                                 error = EWOULDBLOCK;
2078                         }
2079                 }
2080                 if (error)
2081                         break;
2082                 ran_beg = off;  /* for restart */
2083                 error = hammer_ip_next(cursor);
2084         }
2085         if (cursor->node)
2086                 hammer_cache_node(&ip->cache[1], cursor->node);
2087
2088         if (error == EDEADLK) {
2089                 hammer_done_cursor(cursor);
2090                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2091                 if (error == 0)
2092                         goto retry;
2093         }
2094         if (error == ENOENT)
2095                 error = 0;
2096         return(error);
2097 }
2098
2099 /*
2100  * This backend function deletes the specified record on-disk, similar to
2101  * delete_range but for a specific record.  Unlike the exact deletions
2102  * used when deleting a directory entry this function uses an ASOF search 
2103  * like delete_range.
2104  *
2105  * This function may be called with ip->obj_asof set for a slave snapshot,
2106  * so don't use it.  We always delete non-historical records only.
2107  */
2108 static int
2109 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2110                       hammer_btree_leaf_elm_t leaf)
2111 {
2112         hammer_transaction_t trans = cursor->trans;
2113         int error;
2114
2115         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2116 retry:
2117         hammer_normalize_cursor(cursor);
2118         cursor->key_beg = leaf->base;
2119         cursor->asof = HAMMER_MAX_TID;
2120         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2121         cursor->flags |= HAMMER_CURSOR_ASOF;
2122         cursor->flags |= HAMMER_CURSOR_BACKEND;
2123         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2124
2125         error = hammer_btree_lookup(cursor);
2126         if (error == 0) {
2127                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2128         }
2129         if (error == EDEADLK) {
2130                 hammer_done_cursor(cursor);
2131                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2132                 if (error == 0)
2133                         goto retry;
2134         }
2135         return(error);
2136 }
2137
2138 /*
2139  * This function deletes remaining auxillary records when an inode is
2140  * being deleted.  This function explicitly does not delete the
2141  * inode record, directory entry, data, or db records.  Those must be
2142  * properly disposed of prior to this call.
2143  */
2144 int
2145 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2146 {
2147         hammer_transaction_t trans = cursor->trans;
2148         hammer_btree_leaf_elm_t leaf;
2149         int error;
2150
2151         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2152 retry:
2153         hammer_normalize_cursor(cursor);
2154         cursor->key_beg.localization = ip->obj_localization +
2155                                        HAMMER_LOCALIZE_MISC;
2156         cursor->key_beg.obj_id = ip->obj_id;
2157         cursor->key_beg.create_tid = 0;
2158         cursor->key_beg.delete_tid = 0;
2159         cursor->key_beg.obj_type = 0;
2160         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2161         cursor->key_beg.key = HAMMER_MIN_KEY;
2162
2163         cursor->key_end = cursor->key_beg;
2164         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2165         cursor->key_end.key = HAMMER_MAX_KEY;
2166
2167         cursor->asof = ip->obj_asof;
2168         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2169         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2170         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2171         cursor->flags |= HAMMER_CURSOR_BACKEND;
2172
2173         error = hammer_ip_first(cursor);
2174
2175         /*
2176          * Iterate through matching records and mark them as deleted.
2177          */
2178         while (error == 0) {
2179                 leaf = cursor->leaf;
2180
2181                 KKASSERT(leaf->base.delete_tid == 0);
2182
2183                 /*
2184                  * Mark the record and B-Tree entry as deleted.  This will
2185                  * also physically delete the B-Tree entry, record, and
2186                  * data if the retention policy dictates.  The function
2187                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2188                  * to retest the new 'current' element.
2189                  *
2190                  * Directory entries (and delete-on-disk directory entries)
2191                  * must be synced and cannot be deleted.
2192                  */
2193                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2194                 ++*countp;
2195                 if (error)
2196                         break;
2197                 error = hammer_ip_next(cursor);
2198         }
2199         if (cursor->node)
2200                 hammer_cache_node(&ip->cache[1], cursor->node);
2201         if (error == EDEADLK) {
2202                 hammer_done_cursor(cursor);
2203                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2204                 if (error == 0)
2205                         goto retry;
2206         }
2207         if (error == ENOENT)
2208                 error = 0;
2209         return(error);
2210 }
2211
2212 /*
2213  * Delete the record at the current cursor.  On success the cursor will
2214  * be positioned appropriately for an iteration but may no longer be at
2215  * a leaf node.
2216  *
2217  * This routine is only called from the backend.
2218  *
2219  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2220  * cursor and retry.
2221  */
2222 int
2223 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2224                         hammer_tid_t tid)
2225 {
2226         hammer_record_t iprec;
2227         hammer_mount_t hmp;
2228         int error;
2229
2230         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2231         KKASSERT(tid != 0);
2232         hmp = cursor->node->hmp;
2233
2234         /*
2235          * In-memory (unsynchronized) records can simply be freed.  This
2236          * only occurs in range iterations since all other records are
2237          * individually synchronized.  Thus there should be no confusion with
2238          * the interlock.
2239          *
2240          * An in-memory record may be deleted before being committed to disk,
2241          * but could have been accessed in the mean time.  The reservation
2242          * code will deal with the case.
2243          */
2244         if (hammer_cursor_inmem(cursor)) {
2245                 iprec = cursor->iprec;
2246                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2247                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2248                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2249                 KKASSERT(iprec->ip == ip);
2250                 ++ip->rec_generation;
2251                 return(0);
2252         }
2253
2254         /*
2255          * On-disk records are marked as deleted by updating their delete_tid.
2256          * This does not effect their position in the B-Tree (which is based
2257          * on their create_tid).
2258          *
2259          * Frontend B-Tree operations track inodes so we tell 
2260          * hammer_delete_at_cursor() not to.
2261          */
2262         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2263
2264         if (error == 0) {
2265                 error = hammer_delete_at_cursor(
2266                                 cursor,
2267                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2268                                 cursor->trans->tid,
2269                                 cursor->trans->time32,
2270                                 0, NULL);
2271         }
2272         return(error);
2273 }
2274
2275 /*
2276  * Used to write a generic record w/optional data to the media b-tree
2277  * when no inode context is available.  Used by the mirroring and
2278  * snapshot code.
2279  *
2280  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2281  * flagged for backend operation and not flagged ASOF (since we are
2282  * doing an insertion).
2283  *
2284  * This function will acquire the appropriate sync lock and will set
2285  * the cursor insertion flag for the operation, do the btree lookup,
2286  * and the insertion, and clear the insertion flag and sync lock before
2287  * returning.  The cursor state will be such that the caller can continue
2288  * scanning (used by the mirroring code).
2289  *
2290  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2291  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2292  *
2293  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2294  *                retry.
2295  *
2296  *       EALREADY can be returned if the record already exists (WARNING,
2297  *                because ASOF cannot be used no check is made for illegal
2298  *                duplicates).
2299  *
2300  * NOTE: Do not use the function for normal inode-related records as this
2301  *       functions goes directly to the media and is not integrated with
2302  *       in-memory records.
2303  */
2304 int
2305 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2306                         void *udata, int mode)
2307 {
2308         hammer_transaction_t trans;
2309         hammer_buffer_t data_buffer;
2310         hammer_off_t ndata_offset;
2311         hammer_tid_t high_tid;
2312         void *ndata;
2313         int error;
2314         int doprop;
2315
2316         trans = cursor->trans;
2317         data_buffer = NULL;
2318         ndata_offset = 0;
2319         doprop = 0;
2320
2321         KKASSERT((cursor->flags &
2322                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2323                   (HAMMER_CURSOR_BACKEND));
2324
2325         hammer_sync_lock_sh(trans);
2326
2327         if (leaf->data_len) {
2328                 ndata = hammer_alloc_data(trans, leaf->data_len,
2329                                           leaf->base.rec_type,
2330                                           &ndata_offset, &data_buffer,
2331                                           0, &error);
2332                 if (ndata == NULL) {
2333                         hammer_sync_unlock(trans);
2334                         return (error);
2335                 }
2336                 leaf->data_offset = ndata_offset;
2337                 hammer_modify_buffer(trans, data_buffer, NULL, 0);
2338
2339                 switch(mode) {
2340                 case HAMMER_CREATE_MODE_UMIRROR:
2341                         error = copyin(udata, ndata, leaf->data_len);
2342                         if (error == 0) {
2343                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2344                                         kprintf("data crc mismatch on pipe\n");
2345                                         error = EINVAL;
2346                                 } else {
2347                                         error = hammer_cursor_localize_data(
2348                                                         ndata, leaf);
2349                                 }
2350                         }
2351                         break;
2352                 case HAMMER_CREATE_MODE_SYS:
2353                         bcopy(udata, ndata, leaf->data_len);
2354                         error = 0;
2355                         hammer_crc_set_leaf(ndata, leaf);
2356                         break;
2357                 default:
2358                         panic("hammer: hammer_create_at_cursor: bad mode %d",
2359                                 mode);
2360                         break; /* NOT REACHED */
2361                 }
2362                 hammer_modify_buffer_done(data_buffer);
2363         } else {
2364                 leaf->data_offset = 0;
2365                 error = 0;
2366                 ndata = NULL;
2367         }
2368         if (error)
2369                 goto failed;
2370
2371         /*
2372          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2373          */
2374         cursor->flags |= HAMMER_CURSOR_INSERT;
2375         error = hammer_btree_lookup(cursor);
2376         if (error != ENOENT) {
2377                 if (error == 0)
2378                         error = EALREADY;
2379                 goto failed;
2380         }
2381         error = hammer_btree_insert(cursor, leaf, &doprop);
2382
2383         /*
2384          * Cursor is left on current element, we want to skip it now.
2385          * (in case the caller is scanning)
2386          */
2387         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2388         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2389
2390         /*
2391          * If the insertion happens to be creating (and not just replacing)
2392          * an inode we have to track it.
2393          */
2394         if (error == 0 &&
2395             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2396             leaf->base.delete_tid == 0) {
2397                 hammer_modify_volume_field(trans, trans->rootvol,
2398                                            vol0_stat_inodes);
2399                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2400                 hammer_modify_volume_done(trans->rootvol);
2401         }
2402
2403         /*
2404          * vol0_next_tid must track the highest TID stored in the filesystem.
2405          * We do not need to generate undo for this update.
2406          */
2407         high_tid = leaf->base.create_tid;
2408         if (high_tid < leaf->base.delete_tid)
2409                 high_tid = leaf->base.delete_tid;
2410         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2411                 hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2412                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2413                 hammer_modify_volume_done(trans->rootvol);
2414         }
2415
2416         /*
2417          * WARNING!  cursor's leaf pointer may have changed after
2418          *           do_propagation returns.
2419          */
2420         if (error == 0 && doprop)
2421                 hammer_btree_do_propagation(cursor, NULL, leaf);
2422
2423 failed:
2424         /*
2425          * Cleanup
2426          */
2427         if (error && leaf->data_offset) {
2428                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2429
2430         }
2431         hammer_sync_unlock(trans);
2432         if (data_buffer)
2433                 hammer_rel_buffer(data_buffer, 0);
2434         return (error);
2435 }
2436
2437 /*
2438  * Delete the B-Tree element at the current cursor and do any necessary
2439  * mirror propagation.
2440  *
2441  * The cursor must be properly positioned for an iteration on return but
2442  * may be pointing at an internal element.
2443  *
2444  * An element can be un-deleted by passing a delete_tid of 0 with
2445  * HAMMER_DELETE_ADJUST.
2446  */
2447 int
2448 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2449                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2450                         int track, int64_t *stat_bytes)
2451 {
2452         struct hammer_btree_leaf_elm save_leaf;
2453         hammer_transaction_t trans;
2454         hammer_btree_leaf_elm_t leaf;
2455         hammer_node_t node;
2456         hammer_btree_elm_t elm;
2457         hammer_off_t data_offset;
2458         int32_t data_len;
2459         u_int16_t rec_type;
2460         int error;
2461         int icount;
2462         int doprop;
2463
2464         error = hammer_cursor_upgrade(cursor);
2465         if (error)
2466                 return(error);
2467
2468         trans = cursor->trans;
2469         node = cursor->node;
2470         elm = &node->ondisk->elms[cursor->index];
2471         leaf = &elm->leaf;
2472         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2473
2474         hammer_sync_lock_sh(trans);
2475         doprop = 0;
2476         icount = 0;
2477
2478         /*
2479          * Adjust the delete_tid.  Update the mirror_tid propagation field
2480          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2481          */
2482         if (delete_flags & HAMMER_DELETE_ADJUST) {
2483                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2484                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2485                                 icount = -1;
2486                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2487                                 icount = 1;
2488                 }
2489
2490                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2491                 elm->leaf.base.delete_tid = delete_tid;
2492                 elm->leaf.delete_ts = delete_ts;
2493                 hammer_modify_node_done(node);
2494
2495                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2496                         hammer_modify_node_field(trans, node, mirror_tid);
2497                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2498                         hammer_modify_node_done(node);
2499                         doprop = 1;
2500                         if (hammer_debug_general & 0x0002) {
2501                                 kprintf("delete_at_cursor: propagate %016llx"
2502                                         " @%016llx\n",
2503                                         (long long)elm->leaf.base.delete_tid,
2504                                         (long long)node->node_offset);
2505                         }
2506                 }
2507
2508                 /*
2509                  * Adjust for the iteration.  We have deleted the current
2510                  * element and want to clear ATEDISK so the iteration does
2511                  * not skip the element after, which now becomes the current
2512                  * element.  This element must be re-tested if doing an
2513                  * iteration, which is handled by the RETEST flag.
2514                  */
2515                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2516                         cursor->flags |= HAMMER_CURSOR_RETEST;
2517                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2518                 }
2519
2520                 /*
2521                  * An on-disk record cannot have the same delete_tid
2522                  * as its create_tid.  In a chain of record updates
2523                  * this could result in a duplicate record.
2524                  */
2525                 KKASSERT(elm->leaf.base.delete_tid !=
2526                          elm->leaf.base.create_tid);
2527         }
2528
2529         /*
2530          * Destroy the B-Tree element if asked (typically if a nohistory
2531          * file or mount, or when called by the pruning code).
2532          *
2533          * Adjust the ATEDISK flag to properly support iterations.
2534          */
2535         if (delete_flags & HAMMER_DELETE_DESTROY) {
2536                 data_offset = elm->leaf.data_offset;
2537                 data_len = elm->leaf.data_len;
2538                 rec_type = elm->leaf.base.rec_type;
2539                 if (doprop) {
2540                         save_leaf = elm->leaf;
2541                         leaf = &save_leaf;
2542                 }
2543                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2544                     elm->leaf.base.delete_tid == 0) {
2545                         icount = -1;
2546                 }
2547
2548                 error = hammer_btree_delete(cursor);
2549                 if (error == 0) {
2550                         /*
2551                          * The deletion moves the next element (if any) to
2552                          * the current element position.  We must clear
2553                          * ATEDISK so this element is not skipped and we
2554                          * must set RETEST to force any iteration to re-test
2555                          * the element.
2556                          */
2557                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2558                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2559                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2560                         }
2561                 }
2562                 if (error == 0) {
2563                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2564                         case HAMMER_ZONE_LARGE_DATA:
2565                         case HAMMER_ZONE_SMALL_DATA:
2566                         case HAMMER_ZONE_META:
2567                                 hammer_blockmap_free(trans,
2568                                                      data_offset, data_len);
2569                                 break;
2570                         default:
2571                                 break;
2572                         }
2573                 }
2574         }
2575
2576         /*
2577          * Track inode count and next_tid.  This is used by the mirroring
2578          * and PFS code.  icount can be negative, zero, or positive.
2579          */
2580         if (error == 0 && track) {
2581                 if (icount) {
2582                         hammer_modify_volume_field(trans, trans->rootvol,
2583                                                    vol0_stat_inodes);
2584                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2585                         hammer_modify_volume_done(trans->rootvol);
2586                 }
2587                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2588                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2589                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2590                         hammer_modify_volume_done(trans->rootvol);
2591                 }
2592         }
2593
2594         /*
2595          * mirror_tid propagation occurs if the node's mirror_tid had to be
2596          * updated while adjusting the delete_tid.
2597          *
2598          * This occurs when deleting even in nohistory mode, but does not
2599          * occur when pruning an already-deleted node.
2600          *
2601          * cursor->ip is NULL when called from the pruning, mirroring,
2602          * and pfs code.  If non-NULL propagation will be conditionalized
2603          * on whether the PFS is in no-history mode or not.
2604          *
2605          * WARNING: cursor's leaf pointer may have changed after do_propagation
2606          *          returns!
2607          */
2608         if (doprop) {
2609                 if (cursor->ip)
2610                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2611                 else
2612                         hammer_btree_do_propagation(cursor, NULL, leaf);
2613         }
2614         hammer_sync_unlock(trans);
2615         return (error);
2616 }
2617
2618 /*
2619  * Determine whether we can remove a directory.  This routine checks whether
2620  * a directory is empty or not and enforces flush connectivity.
2621  *
2622  * Flush connectivity requires that we block if the target directory is
2623  * currently flushing, otherwise it may not end up in the same flush group.
2624  *
2625  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2626  */
2627 int
2628 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2629 {
2630         struct hammer_cursor cursor;
2631         int error;
2632
2633         /*
2634          * Check directory empty
2635          */
2636         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2637
2638         cursor.key_beg.localization = ip->obj_localization +
2639                                       hammer_dir_localization(ip);
2640         cursor.key_beg.obj_id = ip->obj_id;
2641         cursor.key_beg.create_tid = 0;
2642         cursor.key_beg.delete_tid = 0;
2643         cursor.key_beg.obj_type = 0;
2644         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2645         cursor.key_beg.key = HAMMER_MIN_KEY;
2646
2647         cursor.key_end = cursor.key_beg;
2648         cursor.key_end.rec_type = 0xFFFF;
2649         cursor.key_end.key = HAMMER_MAX_KEY;
2650
2651         cursor.asof = ip->obj_asof;
2652         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2653
2654         error = hammer_ip_first(&cursor);
2655         if (error == ENOENT)
2656                 error = 0;
2657         else if (error == 0)
2658                 error = ENOTEMPTY;
2659         hammer_done_cursor(&cursor);
2660         return(error);
2661 }
2662
2663 /*
2664  * Localize the data payload.  Directory entries may need their
2665  * localization adjusted.
2666  */
2667 static
2668 int
2669 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2670                             hammer_btree_leaf_elm_t leaf)
2671 {
2672         u_int32_t localization;
2673
2674         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2675                 localization = leaf->base.localization &
2676                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2677                 if (data->entry.localization != localization) {
2678                         data->entry.localization = localization;
2679                         hammer_crc_set_leaf(data, leaf);
2680                 }
2681         }
2682         return(0);
2683 }