Pre-2.0 release: Sync with HAMMER 64 - NFS and cross-device link fixes.
[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 1.90.2.2 2008/07/19 04:51:09 dillon Exp $
35  */
36
37 #include "hammer.h"
38
39 static int hammer_mem_add(hammer_record_t record);
40 static int hammer_mem_lookup(hammer_cursor_t cursor);
41 static int hammer_mem_first(hammer_cursor_t cursor);
42 static int hammer_frontend_trunc_callback(hammer_record_t record,
43                                 void *data __unused);
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
48 struct rec_trunc_info {
49         u_int16_t       rec_type;
50         int64_t         trunc_off;
51 };
52
53 /*
54  * Red-black tree support.  Comparison code for insertion.
55  */
56 static int
57 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
58 {
59         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
60                 return(-1);
61         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
62                 return(1);
63
64         if (rec1->leaf.base.key < rec2->leaf.base.key)
65                 return(-1);
66         if (rec1->leaf.base.key > rec2->leaf.base.key)
67                 return(1);
68
69         /*
70          * Never match against an item deleted by the front-end.
71          *
72          * rec1 is greater then rec2 if rec1 is marked deleted.
73          * rec1 is less then rec2 if rec2 is marked deleted.
74          *
75          * Multiple deleted records may be present, do not return 0
76          * if both are marked deleted.
77          */
78         if (rec1->flags & HAMMER_RECF_DELETED_FE)
79                 return(1);
80         if (rec2->flags & HAMMER_RECF_DELETED_FE)
81                 return(-1);
82
83         return(0);
84 }
85
86 /*
87  * Basic record comparison code similar to hammer_btree_cmp().
88  */
89 static int
90 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
91 {
92         if (elm->rec_type < rec->leaf.base.rec_type)
93                 return(-3);
94         if (elm->rec_type > rec->leaf.base.rec_type)
95                 return(3);
96
97         if (elm->key < rec->leaf.base.key)
98                 return(-2);
99         if (elm->key > rec->leaf.base.key)
100                 return(2);
101
102         /*
103          * Never match against an item deleted by the front-end.
104          * elm is less then rec if rec is marked deleted.
105          */
106         if (rec->flags & HAMMER_RECF_DELETED_FE)
107                 return(-1);
108         return(0);
109 }
110
111 /*
112  * Special LOOKUP_INFO to locate an overlapping record.  This used by
113  * the reservation code to implement small-block records (whos keys will
114  * be different depending on data_len, when representing the same base
115  * offset).
116  *
117  * NOTE: The base file offset of a data record is (key - data_len), not (key).
118  */
119 static int
120 hammer_rec_overlap_compare(hammer_btree_leaf_elm_t leaf, hammer_record_t rec)
121 {
122         if (leaf->base.rec_type < rec->leaf.base.rec_type)
123                 return(-3);
124         if (leaf->base.rec_type > rec->leaf.base.rec_type)
125                 return(3);
126
127         /*
128          * Overlap compare
129          */
130         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
131                 /* leaf_end <= rec_beg */
132                 if (leaf->base.key <= rec->leaf.base.key - rec->leaf.data_len)
133                         return(-2);
134                 /* leaf_beg >= rec_end */
135                 if (leaf->base.key - leaf->data_len >= rec->leaf.base.key)
136                         return(2);
137         } else {
138                 if (leaf->base.key < rec->leaf.base.key)
139                         return(-2);
140                 if (leaf->base.key > rec->leaf.base.key)
141                         return(2);
142         }
143
144         /*
145          * Never match against an item deleted by the front-end.
146          * leaf is less then rec if rec is marked deleted.
147          *
148          * We must still return the proper code for the scan to continue
149          * along the correct branches.
150          */
151         if (rec->flags & HAMMER_RECF_DELETED_FE) {
152                 if (leaf->base.key < rec->leaf.base.key)
153                         return(-2);
154                 if (leaf->base.key > rec->leaf.base.key)
155                         return(2);
156                 return(-1);
157         }
158         return(0);
159 }
160
161 /*
162  * RB_SCAN comparison code for hammer_mem_first().  The argument order
163  * is reversed so the comparison result has to be negated.  key_beg and
164  * key_end are both range-inclusive.
165  *
166  * Localized deletions are not cached in-memory.
167  */
168 static
169 int
170 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
171 {
172         hammer_cursor_t cursor = data;
173         int r;
174
175         r = hammer_rec_cmp(&cursor->key_beg, rec);
176         if (r > 1)
177                 return(-1);
178         r = hammer_rec_cmp(&cursor->key_end, rec);
179         if (r < -1)
180                 return(1);
181         return(0);
182 }
183
184 /*
185  * This compare function is used when simply looking up key_beg.
186  */
187 static
188 int
189 hammer_rec_find_cmp(hammer_record_t rec, void *data)
190 {
191         hammer_cursor_t cursor = data;
192         int r;
193
194         r = hammer_rec_cmp(&cursor->key_beg, rec);
195         if (r > 1)
196                 return(-1);
197         if (r < -1)
198                 return(1);
199         return(0);
200 }
201
202 /*
203  * Locate blocks within the truncation range.  Partial blocks do not count.
204  */
205 static
206 int
207 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
208 {
209         struct rec_trunc_info *info = data;
210
211         if (rec->leaf.base.rec_type < info->rec_type)
212                 return(-1);
213         if (rec->leaf.base.rec_type > info->rec_type)
214                 return(1);
215
216         switch(rec->leaf.base.rec_type) {
217         case HAMMER_RECTYPE_DB:
218                 /*
219                  * DB record key is not beyond the truncation point, retain.
220                  */
221                 if (rec->leaf.base.key < info->trunc_off)
222                         return(-1);
223                 break;
224         case HAMMER_RECTYPE_DATA:
225                 /*
226                  * DATA record offset start is not beyond the truncation point,
227                  * retain.
228                  */
229                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
230                         return(-1);
231                 break;
232         default:
233                 panic("hammer_rec_trunc_cmp: unexpected record type");
234         }
235
236         /*
237          * The record start is >= the truncation point, return match,
238          * the record should be destroyed.
239          */
240         return(0);
241 }
242
243 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
244 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
245                     hammer_rec_overlap_compare, hammer_btree_leaf_elm_t);
246
247 /*
248  * Allocate a record for the caller to finish filling in.  The record is
249  * returned referenced.
250  */
251 hammer_record_t
252 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
253 {
254         hammer_record_t record;
255
256         ++hammer_count_records;
257         record = kmalloc(sizeof(*record), M_HAMMER,
258                          M_WAITOK | M_ZERO | M_USE_RESERVE);
259         record->flush_state = HAMMER_FST_IDLE;
260         record->ip = ip;
261         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
262         record->leaf.data_len = data_len;
263         hammer_ref(&record->lock);
264
265         if (data_len) {
266                 record->data = kmalloc(data_len, M_HAMMER, M_WAITOK | M_ZERO);
267                 record->flags |= HAMMER_RECF_ALLOCDATA;
268                 ++hammer_count_record_datas;
269         }
270
271         return (record);
272 }
273
274 void
275 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
276 {
277         while (record->flush_state == HAMMER_FST_FLUSH) {
278                 record->flags |= HAMMER_RECF_WANTED;
279                 tsleep(record, 0, ident, 0);
280         }
281 }
282
283 /*
284  * Called from the backend, hammer_inode.c, after a record has been
285  * flushed to disk.  The record has been exclusively locked by the
286  * caller and interlocked with BE.
287  *
288  * We clean up the state, unlock, and release the record (the record
289  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
290  */
291 void
292 hammer_flush_record_done(hammer_record_t record, int error)
293 {
294         hammer_inode_t target_ip;
295
296         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
297         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
298
299         if (error) {
300                 /*
301                  * An error occured, the backend was unable to sync the
302                  * record to its media.  Leave the record intact.
303                  */
304                 hammer_critical_error(record->ip->hmp, record->ip, error,
305                                       "while flushing record");
306         }
307
308         --record->flush_group->refs;
309         record->flush_group = NULL;
310
311         if (record->flags & HAMMER_RECF_DELETED_BE) {
312                 if ((target_ip = record->target_ip) != NULL) {
313                         TAILQ_REMOVE(&target_ip->target_list, record,
314                                      target_entry);
315                         record->target_ip = NULL;
316                         hammer_test_inode(target_ip);
317                 }
318                 record->flush_state = HAMMER_FST_IDLE;
319         } else {
320                 if (record->target_ip) {
321                         record->flush_state = HAMMER_FST_SETUP;
322                         hammer_test_inode(record->ip);
323                         hammer_test_inode(record->target_ip);
324                 } else {
325                         record->flush_state = HAMMER_FST_IDLE;
326                 }
327         }
328         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
329         if (record->flags & HAMMER_RECF_WANTED) {
330                 record->flags &= ~HAMMER_RECF_WANTED;
331                 wakeup(record);
332         }
333         hammer_rel_mem_record(record);
334 }
335
336 /*
337  * Release a memory record.  Records marked for deletion are immediately
338  * removed from the RB-Tree but otherwise left intact until the last ref
339  * goes away.
340  */
341 void
342 hammer_rel_mem_record(struct hammer_record *record)
343 {
344         hammer_mount_t hmp;
345         hammer_reserve_t resv;
346         hammer_inode_t ip;
347         hammer_inode_t target_ip;
348
349         hammer_unref(&record->lock);
350
351         if (record->lock.refs == 0) {
352                 /*
353                  * Upon release of the last reference wakeup any waiters.
354                  * The record structure may get destroyed so callers will
355                  * loop up and do a relookup.
356                  *
357                  * WARNING!  Record must be removed from RB-TREE before we
358                  * might possibly block.  hammer_test_inode() can block!
359                  */
360                 ip = record->ip;
361                 hmp = ip->hmp;
362
363                 /*
364                  * Upon release of the last reference a record marked deleted
365                  * is destroyed.
366                  */
367                 if (record->flags & HAMMER_RECF_DELETED_FE) {
368                         KKASSERT(ip->lock.refs > 0);
369                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
370
371                         /*
372                          * target_ip may have zero refs, we have to ref it
373                          * to prevent it from being ripped out from under
374                          * us.
375                          */
376                         if ((target_ip = record->target_ip) != NULL) {
377                                 TAILQ_REMOVE(&target_ip->target_list,
378                                              record, target_entry);
379                                 record->target_ip = NULL;
380                                 hammer_ref(&target_ip->lock);
381                         }
382
383                         if (record->flags & HAMMER_RECF_ONRBTREE) {
384                                 RB_REMOVE(hammer_rec_rb_tree,
385                                           &record->ip->rec_tree,
386                                           record);
387                                 KKASSERT(ip->rsv_recs > 0);
388                                 --hmp->rsv_recs;
389                                 --ip->rsv_recs;
390                                 hmp->rsv_databytes -= record->leaf.data_len;
391                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
392
393                                 if (RB_EMPTY(&record->ip->rec_tree)) {
394                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
395                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
396                                         hammer_test_inode(record->ip);
397                                 }
398                         }
399
400                         /*
401                          * We must wait for any direct-IO to complete before
402                          * we can destroy the record.
403                          */
404                         if (record->flags & HAMMER_RECF_DIRECT_IO)
405                                 hammer_io_direct_wait(record);
406
407
408                         /*
409                          * Do this test after removing record from the B-Tree.
410                          */
411                         if (target_ip) {
412                                 hammer_test_inode(target_ip);
413                                 hammer_rel_inode(target_ip, 0);
414                         }
415
416                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
417                                 --hammer_count_record_datas;
418                                 kfree(record->data, M_HAMMER);
419                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
420                         }
421
422                         /*
423                          * Release the reservation.  If the record was not
424                          * committed return the reservation before
425                          * releasing it.
426                          */
427                         if ((resv = record->resv) != NULL) {
428                                 if ((record->flags & HAMMER_RECF_COMMITTED) == 0) {
429                                         hammer_blockmap_reserve_undo(
430                                                 resv,
431                                                 record->leaf.data_offset,
432                                                 record->leaf.data_len);
433                                 }
434                                 hammer_blockmap_reserve_complete(hmp, resv);
435                                 record->resv = NULL;
436                         }
437                         record->data = NULL;
438                         --hammer_count_records;
439                         kfree(record, M_HAMMER);
440                 }
441         }
442 }
443
444 /*
445  * Record visibility depends on whether the record is being accessed by
446  * the backend or the frontend.
447  *
448  * Return non-zero if the record is visible, zero if it isn't or if it is
449  * deleted.
450  */
451 static __inline
452 int
453 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
454 {
455         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
456                 if (record->flags & HAMMER_RECF_DELETED_BE)
457                         return(0);
458         } else {
459                 if (record->flags & HAMMER_RECF_DELETED_FE)
460                         return(0);
461         }
462         return(1);
463 }
464
465 /*
466  * This callback is used as part of the RB_SCAN function for in-memory
467  * records.  We terminate it (return -1) as soon as we get a match.
468  *
469  * This routine is used by frontend code.
470  *
471  * The primary compare code does not account for ASOF lookups.  This
472  * code handles that case as well as a few others.
473  */
474 static
475 int
476 hammer_rec_scan_callback(hammer_record_t rec, void *data)
477 {
478         hammer_cursor_t cursor = data;
479
480         /*
481          * We terminate on success, so this should be NULL on entry.
482          */
483         KKASSERT(cursor->iprec == NULL);
484
485         /*
486          * Skip if the record was marked deleted.
487          */
488         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
489                 return(0);
490
491         /*
492          * Skip if not visible due to our as-of TID
493          */
494         if (cursor->flags & HAMMER_CURSOR_ASOF) {
495                 if (cursor->asof < rec->leaf.base.create_tid)
496                         return(0);
497                 if (rec->leaf.base.delete_tid &&
498                     cursor->asof >= rec->leaf.base.delete_tid) {
499                         return(0);
500                 }
501         }
502
503         /*
504          * ref the record.  The record is protected from backend B-Tree
505          * interactions by virtue of the cursor's IP lock.
506          */
507         hammer_ref(&rec->lock);
508
509         /*
510          * The record may have been deleted while we were blocked.
511          */
512         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
513                 hammer_rel_mem_record(rec);
514                 return(0);
515         }
516
517         /*
518          * Set the matching record and stop the scan.
519          */
520         cursor->iprec = rec;
521         return(-1);
522 }
523
524
525 /*
526  * Lookup an in-memory record given the key specified in the cursor.  Works
527  * just like hammer_btree_lookup() but operates on an inode's in-memory
528  * record list.
529  *
530  * The lookup must fail if the record is marked for deferred deletion.
531  */
532 static
533 int
534 hammer_mem_lookup(hammer_cursor_t cursor)
535 {
536         int error;
537
538         KKASSERT(cursor->ip);
539         if (cursor->iprec) {
540                 hammer_rel_mem_record(cursor->iprec);
541                 cursor->iprec = NULL;
542         }
543         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
544                                    hammer_rec_scan_callback, cursor);
545
546         if (cursor->iprec == NULL)
547                 error = ENOENT;
548         else
549                 error = 0;
550         return(error);
551 }
552
553 /*
554  * hammer_mem_first() - locate the first in-memory record matching the
555  * cursor within the bounds of the key range.
556  */
557 static
558 int
559 hammer_mem_first(hammer_cursor_t cursor)
560 {
561         hammer_inode_t ip;
562
563         ip = cursor->ip;
564         KKASSERT(ip != NULL);
565
566         if (cursor->iprec) {
567                 hammer_rel_mem_record(cursor->iprec);
568                 cursor->iprec = NULL;
569         }
570
571         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
572                                    hammer_rec_scan_callback, cursor);
573
574         /*
575          * Adjust scan.node and keep it linked into the RB-tree so we can
576          * hold the cursor through third party modifications of the RB-tree.
577          */
578         if (cursor->iprec)
579                 return(0);
580         return(ENOENT);
581 }
582
583 /************************************************************************
584  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
585  ************************************************************************
586  *
587  * These functions manipulate in-memory records.  Such records typically
588  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
589  */
590
591 /*
592  * Add a directory entry (dip,ncp) which references inode (ip).
593  *
594  * Note that the low 32 bits of the namekey are set temporarily to create
595  * a unique in-memory record, and may be modified a second time when the
596  * record is synchronized to disk.  In particular, the low 32 bits cannot be
597  * all 0's when synching to disk, which is not handled here.
598  *
599  * NOTE: bytes does not include any terminating \0 on name, and name might
600  * not be terminated.
601  */
602 int
603 hammer_ip_add_directory(struct hammer_transaction *trans,
604                      struct hammer_inode *dip, const char *name, int bytes,
605                      struct hammer_inode *ip)
606 {
607         struct hammer_cursor cursor;
608         hammer_record_t record;
609         int error;
610         int count;
611         u_int32_t iterator;
612
613         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
614         if (++trans->hmp->namekey_iterator == 0)
615                 ++trans->hmp->namekey_iterator;
616
617         record->type = HAMMER_MEM_RECORD_ADD;
618         record->leaf.base.localization = dip->obj_localization +
619                                          HAMMER_LOCALIZE_MISC;
620         record->leaf.base.obj_id = dip->obj_id;
621         record->leaf.base.key = hammer_directory_namekey(name, bytes);
622         record->leaf.base.key += trans->hmp->namekey_iterator;
623         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
624         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
625         record->data->entry.obj_id = ip->obj_id;
626         record->data->entry.localization = ip->obj_localization;
627         bcopy(name, record->data->entry.name, bytes);
628
629         ++ip->ino_data.nlinks;
630         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
631
632         /*
633          * Find an unused namekey.  Both the in-memory record tree and
634          * the B-Tree are checked.  Exact matches also match create_tid
635          * so use an ASOF search to (mostly) ignore it.
636          *
637          * delete-visibility is set so pending deletions do not give us
638          * a false-negative on our ability to use an iterator.
639          */
640         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
641         cursor.key_beg = record->leaf.base;
642         cursor.flags |= HAMMER_CURSOR_ASOF;
643         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
644         cursor.asof = ip->obj_asof;
645
646         count = 0;
647         while (hammer_ip_lookup(&cursor) == 0) {
648                 iterator = (u_int32_t)record->leaf.base.key + 1;
649                 if (iterator == 0)
650                         iterator = 1;
651                 record->leaf.base.key &= ~0xFFFFFFFFLL;
652                 record->leaf.base.key |= iterator;
653                 cursor.key_beg.key = record->leaf.base.key;
654                 if (++count == 1000000000) {
655                         hammer_rel_mem_record(record);
656                         error = ENOSPC;
657                         goto failed;
658                 }
659         }
660
661         /*
662          * The target inode and the directory entry are bound together.
663          */
664         record->target_ip = ip;
665         record->flush_state = HAMMER_FST_SETUP;
666         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
667
668         /*
669          * The inode now has a dependancy and must be taken out of the idle
670          * state.  An inode not in an idle state is given an extra reference.
671          */
672         if (ip->flush_state == HAMMER_FST_IDLE) {
673                 hammer_ref(&ip->lock);
674                 ip->flush_state = HAMMER_FST_SETUP;
675         }
676         error = hammer_mem_add(record);
677         if (error == 0) {
678                 dip->ino_data.mtime = trans->time;
679                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
680         }
681 failed:
682         hammer_done_cursor(&cursor);
683         return(error);
684 }
685
686 /*
687  * Delete the directory entry and update the inode link count.  The
688  * cursor must be seeked to the directory entry record being deleted.
689  *
690  * The related inode should be share-locked by the caller.  The caller is
691  * on the frontend.
692  *
693  * This function can return EDEADLK requiring the caller to terminate
694  * the cursor, any locks, wait on the returned record, and retry.
695  */
696 int
697 hammer_ip_del_directory(struct hammer_transaction *trans,
698                      hammer_cursor_t cursor, struct hammer_inode *dip,
699                      struct hammer_inode *ip)
700 {
701         hammer_record_t record;
702         int error;
703
704         if (hammer_cursor_inmem(cursor)) {
705                 /*
706                  * In-memory (unsynchronized) records can simply be freed.
707                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
708                  * by the backend, we must still avoid races against the
709                  * backend potentially syncing the record to the media. 
710                  *
711                  * We cannot call hammer_ip_delete_record(), that routine may
712                  * only be called from the backend.
713                  */
714                 record = cursor->iprec;
715                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
716                         KKASSERT(cursor->deadlk_rec == NULL);
717                         hammer_ref(&record->lock);
718                         cursor->deadlk_rec = record;
719                         error = EDEADLK;
720                 } else {
721                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
722                         record->flags |= HAMMER_RECF_DELETED_FE;
723                         error = 0;
724                 }
725         } else {
726                 /*
727                  * If the record is on-disk we have to queue the deletion by
728                  * the record's key.  This also causes lookups to skip the
729                  * record.
730                  */
731                 KKASSERT(dip->flags &
732                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
733                 record = hammer_alloc_mem_record(dip, 0);
734                 record->type = HAMMER_MEM_RECORD_DEL;
735                 record->leaf.base = cursor->leaf->base;
736
737                 record->target_ip = ip;
738                 record->flush_state = HAMMER_FST_SETUP;
739                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
740
741                 /*
742                  * The inode now has a dependancy and must be taken out of
743                  * the idle state.  An inode not in an idle state is given
744                  * an extra reference.
745                  */
746                 if (ip->flush_state == HAMMER_FST_IDLE) {
747                         hammer_ref(&ip->lock);
748                         ip->flush_state = HAMMER_FST_SETUP;
749                 }
750
751                 error = hammer_mem_add(record);
752         }
753
754         /*
755          * One less link.  The file may still be open in the OS even after
756          * all links have gone away.
757          *
758          * We have to terminate the cursor before syncing the inode to
759          * avoid deadlocking against ourselves.  XXX this may no longer
760          * be true.
761          *
762          * If nlinks drops to zero and the vnode is inactive (or there is
763          * no vnode), call hammer_inode_unloadable_check() to zonk the
764          * inode.  If we don't do this here the inode will not be destroyed
765          * on-media until we unmount.
766          */
767         if (error == 0) {
768                 --ip->ino_data.nlinks;
769                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
770                 if (ip->ino_data.nlinks == 0 &&
771                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
772                         hammer_done_cursor(cursor);
773                         hammer_inode_unloadable_check(ip, 1);
774                         hammer_flush_inode(ip, 0);
775                 }
776                 dip->ino_data.mtime = trans->time;
777                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
778
779         }
780         return(error);
781 }
782
783 /*
784  * Add a record to an inode.
785  *
786  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
787  * initialize the following additional fields:
788  *
789  * The related inode should be share-locked by the caller.  The caller is
790  * on the frontend.
791  *
792  * record->rec.entry.base.base.key
793  * record->rec.entry.base.base.rec_type
794  * record->rec.entry.base.base.data_len
795  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
796  */
797 int
798 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
799 {
800         hammer_inode_t ip = record->ip;
801         int error;
802
803         KKASSERT(record->leaf.base.localization != 0);
804         record->leaf.base.obj_id = ip->obj_id;
805         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
806         error = hammer_mem_add(record);
807         return(error);
808 }
809
810 /*
811  * Locate a bulk record in-memory.  Bulk records allow disk space to be
812  * reserved so the front-end can flush large data writes without having
813  * to queue the BIO to the flusher.  Only the related record gets queued
814  * to the flusher.
815  */
816 static hammer_record_t
817 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
818 {
819         hammer_record_t record;
820         struct hammer_btree_leaf_elm leaf;
821
822         bzero(&leaf, sizeof(leaf));
823         leaf.base.obj_id = ip->obj_id;
824         leaf.base.key = file_offset + bytes;
825         leaf.base.create_tid = 0;
826         leaf.base.delete_tid = 0;
827         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
828         leaf.base.obj_type = 0;                 /* unused */
829         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
830         leaf.base.localization = ip->obj_localization + HAMMER_LOCALIZE_MISC;
831         leaf.data_len = bytes;
832
833         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
834         if (record)
835                 hammer_ref(&record->lock);
836         return(record);
837 }
838
839 /*
840  * Reserve blockmap space placemarked with an in-memory record.  
841  *
842  * This routine is called by the frontend in order to be able to directly
843  * flush a buffer cache buffer.  The frontend has locked the related buffer
844  * cache buffers and we should be able to manipulate any overlapping
845  * in-memory records.
846  */
847 hammer_record_t
848 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
849                    int *errorp)
850 {
851         hammer_record_t record;
852         hammer_record_t conflict;
853         int zone;
854         int flags;
855
856         /*
857          * Deal with conflicting in-memory records.  We cannot have multiple
858          * in-memory records for the same offset without seriously confusing
859          * the backend, including but not limited to the backend issuing
860          * delete-create-delete sequences and asserting on the delete_tid
861          * being the same as the create_tid.
862          *
863          * If we encounter a record with the backend interlock set we cannot
864          * immediately delete it without confusing the backend.
865          */
866         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
867                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
868                         conflict->flags |= HAMMER_RECF_WANTED;
869                         tsleep(conflict, 0, "hmrrc3", 0);
870                 } else {
871                         conflict->flags |= HAMMER_RECF_DELETED_FE;
872                 }
873                 hammer_rel_mem_record(conflict);
874         }
875
876         /*
877          * Create a record to cover the direct write.  This is called with
878          * the related BIO locked so there should be no possible conflict.
879          *
880          * The backend is responsible for finalizing the space reserved in
881          * this record.
882          *
883          * XXX bytes not aligned, depend on the reservation code to
884          * align the reservation.
885          */
886         record = hammer_alloc_mem_record(ip, 0);
887         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
888                                            HAMMER_ZONE_SMALL_DATA_INDEX;
889         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
890                                                &record->leaf.data_offset,
891                                                errorp);
892         if (record->resv == NULL) {
893                 kprintf("hammer_ip_add_bulk: reservation failed\n");
894                 hammer_rel_mem_record(record);
895                 return(NULL);
896         }
897         record->type = HAMMER_MEM_RECORD_DATA;
898         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
899         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
900         record->leaf.base.obj_id = ip->obj_id;
901         record->leaf.base.key = file_offset + bytes;
902         record->leaf.base.localization = ip->obj_localization +
903                                          HAMMER_LOCALIZE_MISC;
904         record->leaf.data_len = bytes;
905         hammer_crc_set_leaf(data, &record->leaf);
906         flags = record->flags;
907
908         hammer_ref(&record->lock);      /* mem_add eats a reference */
909         *errorp = hammer_mem_add(record);
910         if (*errorp) {
911                 conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
912                 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
913                         *errorp, conflict, file_offset, bytes);
914                 if (conflict)
915                         kprintf("conflict %lld %d\n", conflict->leaf.base.key, conflict->leaf.data_len);
916                 if (conflict)
917                         hammer_rel_mem_record(conflict);
918         }
919         KKASSERT(*errorp == 0);
920         conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
921         if (conflict != record) {
922                 kprintf("conflict mismatch %p %p %08x\n", conflict, record, record->flags);
923                 if (conflict)
924                     kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict->leaf.base.key, conflict->leaf.data_len, record->leaf.base.key, record->leaf.data_len);
925         }
926         KKASSERT(conflict == record);
927         hammer_rel_mem_record(conflict);
928
929         return (record);
930 }
931
932 /*
933  * Frontend truncation code.  Scan in-memory records only.  On-disk records
934  * and records in a flushing state are handled by the backend.  The vnops
935  * setattr code will handle the block containing the truncation point.
936  *
937  * Partial blocks are not deleted.
938  */
939 int
940 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
941 {
942         struct rec_trunc_info info;
943
944         switch(ip->ino_data.obj_type) {
945         case HAMMER_OBJTYPE_REGFILE:
946                 info.rec_type = HAMMER_RECTYPE_DATA;
947                 break;
948         case HAMMER_OBJTYPE_DBFILE:
949                 info.rec_type = HAMMER_RECTYPE_DB;
950                 break;
951         default:
952                 return(EINVAL);
953         }
954         info.trunc_off = file_size;
955         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
956                                    hammer_frontend_trunc_callback, &info);
957         return(0);
958 }
959
960 static int
961 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
962 {
963         if (record->flags & HAMMER_RECF_DELETED_FE)
964                 return(0);
965         if (record->flush_state == HAMMER_FST_FLUSH)
966                 return(0);
967         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
968         hammer_ref(&record->lock);
969         record->flags |= HAMMER_RECF_DELETED_FE;
970         hammer_rel_mem_record(record);
971         return(0);
972 }
973
974 /*
975  * Return 1 if the caller must check for and delete existing records
976  * before writing out a new data record.
977  *
978  * Return 0 if the caller can just insert the record into the B-Tree without
979  * checking.
980  */
981 static int
982 hammer_record_needs_overwrite_delete(hammer_record_t record)
983 {
984         hammer_inode_t ip = record->ip;
985         int64_t file_offset;
986         int r;
987
988         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
989                 file_offset = record->leaf.base.key;
990         else
991                 file_offset = record->leaf.base.key - record->leaf.data_len;
992         r = (file_offset < ip->save_trunc_off);
993         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
994                 if (ip->save_trunc_off <= record->leaf.base.key)
995                         ip->save_trunc_off = record->leaf.base.key + 1;
996         } else {
997                 if (ip->save_trunc_off < record->leaf.base.key)
998                         ip->save_trunc_off = record->leaf.base.key;
999         }
1000         return(r);
1001 }
1002
1003 /*
1004  * Backend code.  Sync a record to the media.
1005  */
1006 int
1007 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1008 {
1009         hammer_transaction_t trans = cursor->trans;
1010         int64_t file_offset;
1011         int bytes;
1012         void *bdata;
1013         int error;
1014         int doprop;
1015
1016         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1017         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1018         KKASSERT(record->leaf.base.localization != 0);
1019
1020         /*
1021          * If this is a bulk-data record placemarker there may be an existing
1022          * record on-disk, indicating a data overwrite.  If there is the
1023          * on-disk record must be deleted before we can insert our new record.
1024          *
1025          * We've synthesized this record and do not know what the create_tid
1026          * on-disk is, nor how much data it represents.
1027          *
1028          * Keep in mind that (key) for data records is (base_offset + len),
1029          * not (base_offset).  Also, we only want to get rid of on-disk
1030          * records since we are trying to sync our in-memory record, call
1031          * hammer_ip_delete_range() with truncating set to 1 to make sure
1032          * it skips in-memory records.
1033          *
1034          * It is ok for the lookup to return ENOENT.
1035          *
1036          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1037          * to call hammer_ip_delete_range() or not.  This also means we must
1038          * update sync_trunc_off() as we write.
1039          */
1040         if (record->type == HAMMER_MEM_RECORD_DATA &&
1041             hammer_record_needs_overwrite_delete(record)) {
1042                 file_offset = record->leaf.base.key - record->leaf.data_len;
1043                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) & 
1044                         ~HAMMER_BUFMASK;
1045                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1046                 error = hammer_ip_delete_range(
1047                                 cursor, record->ip,
1048                                 file_offset, file_offset + bytes - 1,
1049                                 1);
1050                 if (error && error != ENOENT)
1051                         goto done;
1052         }
1053
1054         /*
1055          * If this is a general record there may be an on-disk version
1056          * that must be deleted before we can insert the new record.
1057          */
1058         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1059                 error = hammer_delete_general(cursor, record->ip,
1060                                               &record->leaf);
1061                 if (error && error != ENOENT)
1062                         goto done;
1063         }
1064
1065         /*
1066          * Setup the cursor.
1067          */
1068         hammer_normalize_cursor(cursor);
1069         cursor->key_beg = record->leaf.base;
1070         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1071         cursor->flags |= HAMMER_CURSOR_BACKEND;
1072         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1073
1074         /*
1075          * Records can wind up on-media before the inode itself is on-media.
1076          * Flag the case.
1077          */
1078         record->ip->flags |= HAMMER_INODE_DONDISK;
1079
1080         /*
1081          * If we are deleting a directory entry an exact match must be
1082          * found on-disk.
1083          */
1084         if (record->type == HAMMER_MEM_RECORD_DEL) {
1085                 error = hammer_btree_lookup(cursor);
1086                 if (error == 0) {
1087                         /* XXX iprec? */
1088                         error = hammer_ip_delete_record(cursor, record->ip,
1089                                                         trans->tid);
1090                         if (error == 0) {
1091                                 record->flags |= HAMMER_RECF_DELETED_FE;
1092                                 record->flags |= HAMMER_RECF_DELETED_BE;
1093                                 record->flags |= HAMMER_RECF_COMMITTED;
1094                         }
1095                 }
1096                 goto done;
1097         }
1098
1099         /*
1100          * We are inserting.
1101          *
1102          * Issue a lookup to position the cursor and locate the cluster.  The
1103          * target key should not exist.  If we are creating a directory entry
1104          * we may have to iterate the low 32 bits of the key to find an unused
1105          * key.
1106          */
1107         hammer_sync_lock_sh(trans);
1108         cursor->flags |= HAMMER_CURSOR_INSERT;
1109         error = hammer_btree_lookup(cursor);
1110         if (hammer_debug_inode)
1111                 kprintf("DOINSERT LOOKUP %d\n", error);
1112         if (error == 0) {
1113                 kprintf("hammer_ip_sync_record: duplicate rec "
1114                         "at (%016llx)\n", record->leaf.base.key);
1115                 Debugger("duplicate record1");
1116                 error = EIO;
1117         }
1118 #if 0
1119         if (record->type == HAMMER_MEM_RECORD_DATA)
1120                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1121                         record->leaf.base.key - record->leaf.data_len,
1122                         record->leaf.data_offset, error);
1123 #endif
1124
1125         if (error != ENOENT)
1126                 goto done_unlock;
1127
1128         /*
1129          * Allocate the record and data.  The result buffers will be
1130          * marked as being modified and further calls to
1131          * hammer_modify_buffer() will result in unneeded UNDO records.
1132          *
1133          * Support zero-fill records (data == NULL and data_len != 0)
1134          */
1135         if (record->type == HAMMER_MEM_RECORD_DATA) {
1136                 /*
1137                  * The data portion of a bulk-data record has already been
1138                  * committed to disk, we need only adjust the layer2
1139                  * statistics in the same transaction as our B-Tree insert.
1140                  */
1141                 KKASSERT(record->leaf.data_offset != 0);
1142                 error = hammer_blockmap_finalize(trans,
1143                                                  record->leaf.data_offset,
1144                                                  record->leaf.data_len);
1145         } else if (record->data && record->leaf.data_len) {
1146                 /*
1147                  * Wholely cached record, with data.  Allocate the data.
1148                  */
1149                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1150                                           record->leaf.base.rec_type,
1151                                           &record->leaf.data_offset,
1152                                           &cursor->data_buffer, &error);
1153                 if (bdata == NULL)
1154                         goto done_unlock;
1155                 hammer_crc_set_leaf(record->data, &record->leaf);
1156                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1157                 bcopy(record->data, bdata, record->leaf.data_len);
1158                 hammer_modify_buffer_done(cursor->data_buffer);
1159         } else {
1160                 /*
1161                  * Wholely cached record, without data.
1162                  */
1163                 record->leaf.data_offset = 0;
1164                 record->leaf.data_crc = 0;
1165         }
1166
1167         /*
1168          * If the record's data was direct-written we cannot insert
1169          * it until the direct-IO has completed.
1170          */
1171         if (record->flags & HAMMER_RECF_DIRECT_IO)
1172                 hammer_io_direct_wait(record);
1173
1174         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1175         if (hammer_debug_inode && error)
1176                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1177
1178         /*
1179          * Our record is on-disk, normally mark the in-memory version as
1180          * deleted.  If the record represented a directory deletion but
1181          * we had to sync a valid directory entry to disk we must convert
1182          * the record to a covering delete so the frontend does not have
1183          * visibility on the synced entry.
1184          */
1185         if (error == 0) {
1186                 if (doprop) {
1187                         hammer_btree_do_propagation(cursor,
1188                                                     record->ip->pfsm,
1189                                                     &record->leaf);
1190                 }
1191                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1192                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1193                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1194                         record->type = HAMMER_MEM_RECORD_DEL;
1195                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1196                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1197                         /* hammer_flush_record_done takes care of the rest */
1198                 } else {
1199                         record->flags |= HAMMER_RECF_DELETED_FE;
1200                         record->flags |= HAMMER_RECF_DELETED_BE;
1201                 }
1202                 record->flags |= HAMMER_RECF_COMMITTED;
1203         } else {
1204                 if (record->leaf.data_offset) {
1205                         hammer_blockmap_free(trans, record->leaf.data_offset,
1206                                              record->leaf.data_len);
1207                 }
1208         }
1209 done_unlock:
1210         hammer_sync_unlock(trans);
1211 done:
1212         return(error);
1213 }
1214
1215 /*
1216  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1217  * entry's key is used to deal with hash collisions in the upper 32 bits.
1218  * A unique 64 bit key is generated in-memory and may be regenerated a
1219  * second time when the directory record is flushed to the on-disk B-Tree.
1220  *
1221  * A referenced record is passed to this function.  This function
1222  * eats the reference.  If an error occurs the record will be deleted.
1223  *
1224  * A copy of the temporary record->data pointer provided by the caller
1225  * will be made.
1226  */
1227 static
1228 int
1229 hammer_mem_add(hammer_record_t record)
1230 {
1231         hammer_mount_t hmp = record->ip->hmp;
1232
1233         /*
1234          * Make a private copy of record->data
1235          */
1236         if (record->data)
1237                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1238
1239         /*
1240          * Insert into the RB tree.  A unique key should have already
1241          * been selected if this is a directory entry.
1242          */
1243         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1244                 record->flags |= HAMMER_RECF_DELETED_FE;
1245                 hammer_rel_mem_record(record);
1246                 return (EEXIST);
1247         }
1248         ++hmp->count_newrecords;
1249         ++hmp->rsv_recs;
1250         ++record->ip->rsv_recs;
1251         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1252         record->flags |= HAMMER_RECF_ONRBTREE;
1253         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1254         hammer_rel_mem_record(record);
1255         return(0);
1256 }
1257
1258 /************************************************************************
1259  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1260  ************************************************************************
1261  *
1262  * These functions augment the B-Tree scanning functions in hammer_btree.c
1263  * by merging in-memory records with on-disk records.
1264  */
1265
1266 /*
1267  * Locate a particular record either in-memory or on-disk.
1268  *
1269  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1270  * NOT be called to iterate results.
1271  */
1272 int
1273 hammer_ip_lookup(hammer_cursor_t cursor)
1274 {
1275         int error;
1276
1277         /*
1278          * If the element is in-memory return it without searching the
1279          * on-disk B-Tree
1280          */
1281         KKASSERT(cursor->ip);
1282         error = hammer_mem_lookup(cursor);
1283         if (error == 0) {
1284                 cursor->leaf = &cursor->iprec->leaf;
1285                 return(error);
1286         }
1287         if (error != ENOENT)
1288                 return(error);
1289
1290         /*
1291          * If the inode has on-disk components search the on-disk B-Tree.
1292          */
1293         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1294                 return(error);
1295         error = hammer_btree_lookup(cursor);
1296         if (error == 0)
1297                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1298         return(error);
1299 }
1300
1301 /*
1302  * Locate the first record within the cursor's key_beg/key_end range,
1303  * restricted to a particular inode.  0 is returned on success, ENOENT
1304  * if no records matched the requested range, or some other error.
1305  *
1306  * When 0 is returned hammer_ip_next() may be used to iterate additional
1307  * records within the requested range.
1308  *
1309  * This function can return EDEADLK, requiring the caller to terminate
1310  * the cursor and try again.
1311  */
1312 int
1313 hammer_ip_first(hammer_cursor_t cursor)
1314 {
1315         hammer_inode_t ip = cursor->ip;
1316         int error;
1317
1318         KKASSERT(ip != NULL);
1319
1320         /*
1321          * Clean up fields and setup for merged scan
1322          */
1323         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1324         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1325         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1326         if (cursor->iprec) {
1327                 hammer_rel_mem_record(cursor->iprec);
1328                 cursor->iprec = NULL;
1329         }
1330
1331         /*
1332          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1333          * exact lookup so if we get ENOENT we have to call the iterate
1334          * function to validate the first record after the begin key.
1335          *
1336          * The ATEDISK flag is used by hammer_btree_iterate to determine
1337          * whether it must index forwards or not.  It is also used here
1338          * to select the next record from in-memory or on-disk.
1339          *
1340          * EDEADLK can only occur if the lookup hit an empty internal
1341          * element and couldn't delete it.  Since this could only occur
1342          * in-range, we can just iterate from the failure point.
1343          */
1344         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1345                 error = hammer_btree_lookup(cursor);
1346                 if (error == ENOENT || error == EDEADLK) {
1347                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1348                         if (hammer_debug_general & 0x2000)
1349                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1350                         error = hammer_btree_iterate(cursor);
1351                 }
1352                 if (error && error != ENOENT) 
1353                         return(error);
1354                 if (error == 0) {
1355                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1356                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1357                 } else {
1358                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1359                 }
1360         }
1361
1362         /*
1363          * Search the in-memory record list (Red-Black tree).  Unlike the
1364          * B-Tree search, mem_first checks for records in the range.
1365          */
1366         error = hammer_mem_first(cursor);
1367         if (error && error != ENOENT)
1368                 return(error);
1369         if (error == 0) {
1370                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1371                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1372                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1373                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1374         }
1375
1376         /*
1377          * This will return the first matching record.
1378          */
1379         return(hammer_ip_next(cursor));
1380 }
1381
1382 /*
1383  * Retrieve the next record in a merged iteration within the bounds of the
1384  * cursor.  This call may be made multiple times after the cursor has been
1385  * initially searched with hammer_ip_first().
1386  *
1387  * 0 is returned on success, ENOENT if no further records match the
1388  * requested range, or some other error code is returned.
1389  */
1390 int
1391 hammer_ip_next(hammer_cursor_t cursor)
1392 {
1393         hammer_btree_elm_t elm;
1394         hammer_record_t rec, save;
1395         int error;
1396         int r;
1397
1398 next_btree:
1399         /*
1400          * Load the current on-disk and in-memory record.  If we ate any
1401          * records we have to get the next one. 
1402          *
1403          * If we deleted the last on-disk record we had scanned ATEDISK will
1404          * be clear and DELBTREE will be set, forcing a call to iterate. The
1405          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1406          * element.  If ATEDISK is set, iterate will skip the 'current'
1407          * element.
1408          *
1409          * Get the next on-disk record
1410          */
1411         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1412                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1413                         error = hammer_btree_iterate(cursor);
1414                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1415                         if (error == 0) {
1416                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1417                                 hammer_cache_node(&cursor->ip->cache[1],
1418                                                   cursor->node);
1419                         } else {
1420                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1421                                                  HAMMER_CURSOR_ATEDISK;
1422                         }
1423                 }
1424         }
1425
1426 next_memory:
1427         /*
1428          * Get the next in-memory record.
1429          *
1430          * hammer_rec_scan_cmp:  Is the record still in our general range,
1431          *                       (non-inclusive of snapshot exclusions)?
1432          * hammer_rec_scan_callback: Is the record in our snapshot?
1433          */
1434         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1435                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1436                         save = cursor->iprec;
1437                         cursor->iprec = NULL;
1438                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1439                         while (rec) {
1440                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1441                                         break;
1442                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1443                                         break;
1444                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1445                         }
1446                         if (save)
1447                                 hammer_rel_mem_record(save);
1448                         if (cursor->iprec) {
1449                                 KKASSERT(cursor->iprec == rec);
1450                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1451                         } else {
1452                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1453                         }
1454                 }
1455         }
1456
1457         /*
1458          * The memory record may have become stale while being held in
1459          * cursor->iprec.  We are interlocked against the backend on 
1460          * with regards to B-Tree entries.
1461          */
1462         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1463                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1464                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1465                         goto next_memory;
1466                 }
1467         }
1468
1469         /*
1470          * Extract either the disk or memory record depending on their
1471          * relative position.
1472          */
1473         error = 0;
1474         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1475         case 0:
1476                 /*
1477                  * Both entries valid.   Compare the entries and nominally
1478                  * return the first one in the sort order.  Numerous cases
1479                  * require special attention, however.
1480                  */
1481                 elm = &cursor->node->ondisk->elms[cursor->index];
1482                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1483
1484                 /*
1485                  * If the two entries differ only by their key (-2/2) or
1486                  * create_tid (-1/1), and are DATA records, we may have a
1487                  * nominal match.  We have to calculate the base file
1488                  * offset of the data.
1489                  */
1490                 if (r <= 2 && r >= -2 && r != 0 &&
1491                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1492                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1493                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1494                         int64_t base2 = cursor->iprec->leaf.base.key -
1495                                         cursor->iprec->leaf.data_len;
1496                         if (base1 == base2)
1497                                 r = 0;
1498                 }
1499
1500                 if (r < 0) {
1501                         error = hammer_btree_extract(cursor,
1502                                                      HAMMER_CURSOR_GET_LEAF);
1503                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1504                         break;
1505                 }
1506
1507                 /*
1508                  * If the entries match exactly the memory entry is either
1509                  * an on-disk directory entry deletion or a bulk data
1510                  * overwrite.  If it is a directory entry deletion we eat
1511                  * both entries.
1512                  *
1513                  * For the bulk-data overwrite case it is possible to have
1514                  * visibility into both, which simply means the syncer
1515                  * hasn't gotten around to doing the delete+insert sequence
1516                  * on the B-Tree.  Use the memory entry and throw away the
1517                  * on-disk entry.
1518                  *
1519                  * If the in-memory record is not either of these we
1520                  * probably caught the syncer while it was syncing it to
1521                  * the media.  Since we hold a shared lock on the cursor,
1522                  * the in-memory record had better be marked deleted at
1523                  * this point.
1524                  */
1525                 if (r == 0) {
1526                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1527                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1528                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1529                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1530                                         goto next_btree;
1531                                 }
1532                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1533                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1534                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1535                                 }
1536                                 /* fall through to memory entry */
1537                         } else {
1538                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1539                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1540                                 goto next_memory;
1541                         }
1542                 }
1543                 /* fall through to the memory entry */
1544         case HAMMER_CURSOR_ATEDISK:
1545                 /*
1546                  * Only the memory entry is valid.
1547                  */
1548                 cursor->leaf = &cursor->iprec->leaf;
1549                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1550
1551                 /*
1552                  * If the memory entry is an on-disk deletion we should have
1553                  * also had found a B-Tree record.  If the backend beat us
1554                  * to it it would have interlocked the cursor and we should
1555                  * have seen the in-memory record marked DELETED_FE.
1556                  */
1557                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1558                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1559                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1560                 }
1561                 break;
1562         case HAMMER_CURSOR_ATEMEM:
1563                 /*
1564                  * Only the disk entry is valid
1565                  */
1566                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1567                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1568                 break;
1569         default:
1570                 /*
1571                  * Neither entry is valid
1572                  *
1573                  * XXX error not set properly
1574                  */
1575                 cursor->leaf = NULL;
1576                 error = ENOENT;
1577                 break;
1578         }
1579         return(error);
1580 }
1581
1582 /*
1583  * Resolve the cursor->data pointer for the current cursor position in
1584  * a merged iteration.
1585  */
1586 int
1587 hammer_ip_resolve_data(hammer_cursor_t cursor)
1588 {
1589         hammer_record_t record;
1590         int error;
1591
1592         if (hammer_cursor_inmem(cursor)) {
1593                 /*
1594                  * The data associated with an in-memory record is usually
1595                  * kmalloced, but reserve-ahead data records will have an
1596                  * on-disk reference.
1597                  *
1598                  * NOTE: Reserve-ahead data records must be handled in the
1599                  * context of the related high level buffer cache buffer
1600                  * to interlock against async writes.
1601                  */
1602                 record = cursor->iprec;
1603                 cursor->data = record->data;
1604                 error = 0;
1605                 if (cursor->data == NULL) {
1606                         KKASSERT(record->leaf.base.rec_type ==
1607                                  HAMMER_RECTYPE_DATA);
1608                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1609                                                     record->leaf.data_offset,
1610                                                     record->leaf.data_len,
1611                                                     &error,
1612                                                     &cursor->data_buffer);
1613                 }
1614         } else {
1615                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1616                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1617         }
1618         return(error);
1619 }
1620
1621 /*
1622  * Backend truncation / record replacement - delete records in range.
1623  *
1624  * Delete all records within the specified range for inode ip.  In-memory
1625  * records still associated with the frontend are ignored. 
1626  *
1627  * If truncating is non-zero in-memory records associated with the back-end
1628  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1629  *
1630  * NOTES:
1631  *
1632  *      * An unaligned range will cause new records to be added to cover
1633  *        the edge cases. (XXX not implemented yet).
1634  *
1635  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1636  *        also do not deal with unaligned ranges.
1637  *
1638  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1639  *
1640  *      * Record keys for regular file data have to be special-cased since
1641  *        they indicate the end of the range (key = base + bytes).
1642  *
1643  *      * This function may be asked to delete ridiculously huge ranges, for
1644  *        example if someone truncates or removes a 1TB regular file.  We
1645  *        must be very careful on restarts and we may have to stop w/
1646  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1647  */
1648 int
1649 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1650                        int64_t ran_beg, int64_t ran_end, int truncating)
1651 {
1652         hammer_transaction_t trans = cursor->trans;
1653         hammer_btree_leaf_elm_t leaf;
1654         int error;
1655         int64_t off;
1656         int64_t tmp64;
1657
1658 #if 0
1659         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1660 #endif
1661
1662         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1663 retry:
1664         hammer_normalize_cursor(cursor);
1665         cursor->key_beg.localization = ip->obj_localization +
1666                                        HAMMER_LOCALIZE_MISC;
1667         cursor->key_beg.obj_id = ip->obj_id;
1668         cursor->key_beg.create_tid = 0;
1669         cursor->key_beg.delete_tid = 0;
1670         cursor->key_beg.obj_type = 0;
1671
1672         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1673                 cursor->key_beg.key = ran_beg;
1674                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1675         } else {
1676                 /*
1677                  * The key in the B-Tree is (base+bytes), so the first possible
1678                  * matching key is ran_beg + 1.
1679                  */
1680                 cursor->key_beg.key = ran_beg + 1;
1681                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1682         }
1683
1684         cursor->key_end = cursor->key_beg;
1685         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1686                 cursor->key_end.key = ran_end;
1687         } else {
1688                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1689                 if (tmp64 < ran_end)
1690                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1691                 else
1692                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1693         }
1694
1695         cursor->asof = ip->obj_asof;
1696         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1697         cursor->flags |= HAMMER_CURSOR_ASOF;
1698         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1699         cursor->flags |= HAMMER_CURSOR_BACKEND;
1700         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1701
1702         error = hammer_ip_first(cursor);
1703
1704         /*
1705          * Iterate through matching records and mark them as deleted.
1706          */
1707         while (error == 0) {
1708                 leaf = cursor->leaf;
1709
1710                 KKASSERT(leaf->base.delete_tid == 0);
1711                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1712
1713                 /*
1714                  * There may be overlap cases for regular file data.  Also
1715                  * remember the key for a regular file record is (base + len),
1716                  * NOT (base).
1717                  *
1718                  * Note that do to duplicates (mem & media) allowed by
1719                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1720                  */
1721                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1722                         off = leaf->base.key - leaf->data_len;
1723                         /*
1724                          * Check the left edge case.  We currently do not
1725                          * split existing records.
1726                          */
1727                         if (off < ran_beg && leaf->base.key > ran_beg) {
1728                                 panic("hammer left edge case %016llx %d\n",
1729                                         leaf->base.key, leaf->data_len);
1730                         }
1731
1732                         /*
1733                          * Check the right edge case.  Note that the
1734                          * record can be completely out of bounds, which
1735                          * terminates the search.
1736                          *
1737                          * base->key is exclusive of the right edge while
1738                          * ran_end is inclusive of the right edge.  The
1739                          * (key - data_len) left boundary is inclusive.
1740                          *
1741                          * XXX theory-check this test at some point, are
1742                          * we missing a + 1 somewhere?  Note that ran_end
1743                          * could overflow.
1744                          */
1745                         if (leaf->base.key - 1 > ran_end) {
1746                                 if (leaf->base.key - leaf->data_len > ran_end)
1747                                         break;
1748                                 panic("hammer right edge case\n");
1749                         }
1750                 } else {
1751                         off = leaf->base.key;
1752                 }
1753
1754                 /*
1755                  * Delete the record.  When truncating we do not delete
1756                  * in-memory (data) records because they represent data
1757                  * written after the truncation.
1758                  *
1759                  * This will also physically destroy the B-Tree entry and
1760                  * data if the retention policy dictates.  The function
1761                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1762                  * uses to perform a fixup.
1763                  */
1764                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1765                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1766                         /*
1767                          * If we have built up too many meta-buffers we risk
1768                          * deadlocking the kernel and must stop.  This can
1769                          * occur when deleting ridiculously huge files.
1770                          * sync_trunc_off is updated so the next cycle does
1771                          * not re-iterate records we have already deleted.
1772                          *
1773                          * This is only done with formal truncations.
1774                          */
1775                         if (truncating > 1 && error == 0 &&
1776                             hammer_flusher_meta_limit(ip->hmp)) {
1777                                 ip->sync_trunc_off = off;
1778                                 error = EWOULDBLOCK;
1779                         }
1780                 }
1781                 if (error)
1782                         break;
1783                 ran_beg = off;  /* for restart */
1784                 error = hammer_ip_next(cursor);
1785         }
1786         if (cursor->node)
1787                 hammer_cache_node(&ip->cache[1], cursor->node);
1788
1789         if (error == EDEADLK) {
1790                 hammer_done_cursor(cursor);
1791                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1792                 if (error == 0)
1793                         goto retry;
1794         }
1795         if (error == ENOENT)
1796                 error = 0;
1797         return(error);
1798 }
1799
1800 /*
1801  * This backend function deletes the specified record on-disk, similar to
1802  * delete_range but for a specific record.  Unlike the exact deletions
1803  * used when deleting a directory entry this function uses an ASOF search 
1804  * like delete_range.
1805  *
1806  * This function may be called with ip->obj_asof set for a slave snapshot,
1807  * so don't use it.  We always delete non-historical records only.
1808  */
1809 static int
1810 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
1811                       hammer_btree_leaf_elm_t leaf)
1812 {
1813         hammer_transaction_t trans = cursor->trans;
1814         int error;
1815
1816         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1817 retry:
1818         hammer_normalize_cursor(cursor);
1819         cursor->key_beg = leaf->base;
1820         cursor->asof = HAMMER_MAX_TID;
1821         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1822         cursor->flags |= HAMMER_CURSOR_ASOF;
1823         cursor->flags |= HAMMER_CURSOR_BACKEND;
1824         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1825
1826         error = hammer_btree_lookup(cursor);
1827         if (error == 0) {
1828                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1829         }
1830         if (error == EDEADLK) {
1831                 hammer_done_cursor(cursor);
1832                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1833                 if (error == 0)
1834                         goto retry;
1835         }
1836         return(error);
1837 }
1838
1839 /*
1840  * This function deletes remaining auxillary records when an inode is
1841  * being deleted.  This function explicitly does not delete the
1842  * inode record, directory entry, data, or db records.  Those must be
1843  * properly disposed of prior to this call.
1844  */
1845 int
1846 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
1847 {
1848         hammer_transaction_t trans = cursor->trans;
1849         hammer_btree_leaf_elm_t leaf;
1850         int error;
1851
1852         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1853 retry:
1854         hammer_normalize_cursor(cursor);
1855         cursor->key_beg.localization = ip->obj_localization +
1856                                        HAMMER_LOCALIZE_MISC;
1857         cursor->key_beg.obj_id = ip->obj_id;
1858         cursor->key_beg.create_tid = 0;
1859         cursor->key_beg.delete_tid = 0;
1860         cursor->key_beg.obj_type = 0;
1861         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
1862         cursor->key_beg.key = HAMMER_MIN_KEY;
1863
1864         cursor->key_end = cursor->key_beg;
1865         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
1866         cursor->key_end.key = HAMMER_MAX_KEY;
1867
1868         cursor->asof = ip->obj_asof;
1869         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1870         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1871         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1872         cursor->flags |= HAMMER_CURSOR_BACKEND;
1873
1874         error = hammer_ip_first(cursor);
1875
1876         /*
1877          * Iterate through matching records and mark them as deleted.
1878          */
1879         while (error == 0) {
1880                 leaf = cursor->leaf;
1881
1882                 KKASSERT(leaf->base.delete_tid == 0);
1883
1884                 /*
1885                  * Mark the record and B-Tree entry as deleted.  This will
1886                  * also physically delete the B-Tree entry, record, and
1887                  * data if the retention policy dictates.  The function
1888                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1889                  * uses to perform a fixup.
1890                  *
1891                  * Directory entries (and delete-on-disk directory entries)
1892                  * must be synced and cannot be deleted.
1893                  */
1894                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1895                 ++*countp;
1896                 if (error)
1897                         break;
1898                 error = hammer_ip_next(cursor);
1899         }
1900         if (cursor->node)
1901                 hammer_cache_node(&ip->cache[1], cursor->node);
1902         if (error == EDEADLK) {
1903                 hammer_done_cursor(cursor);
1904                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1905                 if (error == 0)
1906                         goto retry;
1907         }
1908         if (error == ENOENT)
1909                 error = 0;
1910         return(error);
1911 }
1912
1913 /*
1914  * Delete the record at the current cursor.  On success the cursor will
1915  * be positioned appropriately for an iteration but may no longer be at
1916  * a leaf node.
1917  *
1918  * This routine is only called from the backend.
1919  *
1920  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1921  * cursor and retry.
1922  */
1923 int
1924 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1925                         hammer_tid_t tid)
1926 {
1927         hammer_record_t iprec;
1928         hammer_btree_elm_t elm;
1929         hammer_mount_t hmp;
1930         int error;
1931
1932         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1933         KKASSERT(tid != 0);
1934         hmp = cursor->node->hmp;
1935
1936         /*
1937          * In-memory (unsynchronized) records can simply be freed.  This
1938          * only occurs in range iterations since all other records are
1939          * individually synchronized.  Thus there should be no confusion with
1940          * the interlock.
1941          *
1942          * An in-memory record may be deleted before being committed to disk,
1943          * but could have been accessed in the mean time.  The reservation
1944          * code will deal with the case.
1945          */
1946         if (hammer_cursor_inmem(cursor)) {
1947                 iprec = cursor->iprec;
1948                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1949                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1950                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1951                 return(0);
1952         }
1953
1954         /*
1955          * On-disk records are marked as deleted by updating their delete_tid.
1956          * This does not effect their position in the B-Tree (which is based
1957          * on their create_tid).
1958          *
1959          * Frontend B-Tree operations track inodes so we tell 
1960          * hammer_delete_at_cursor() not to.
1961          */
1962         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1963         elm = NULL;
1964
1965         if (error == 0) {
1966                 error = hammer_delete_at_cursor(
1967                                 cursor,
1968                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
1969                                 cursor->trans->tid,
1970                                 cursor->trans->time32,
1971                                 0, NULL);
1972         }
1973         return(error);
1974 }
1975
1976 /*
1977  * Delete the B-Tree element at the current cursor and do any necessary
1978  * mirror propagation.
1979  *
1980  * The cursor must be properly positioned for an iteration on return but
1981  * may be pointing at an internal element.
1982  *
1983  * An element can be un-deleted by passing a delete_tid of 0 with
1984  * HAMMER_DELETE_ADJUST.
1985  */
1986 int
1987 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1988                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1989                         int track, int64_t *stat_bytes)
1990 {
1991         struct hammer_btree_leaf_elm save_leaf;
1992         hammer_transaction_t trans;
1993         hammer_btree_leaf_elm_t leaf;
1994         hammer_node_t node;
1995         hammer_btree_elm_t elm;
1996         hammer_off_t data_offset;
1997         int32_t data_len;
1998         u_int16_t rec_type;
1999         int error;
2000         int icount;
2001         int doprop;
2002
2003         error = hammer_cursor_upgrade(cursor);
2004         if (error)
2005                 return(error);
2006
2007         trans = cursor->trans;
2008         node = cursor->node;
2009         elm = &node->ondisk->elms[cursor->index];
2010         leaf = &elm->leaf;
2011         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2012
2013         hammer_sync_lock_sh(trans);
2014         doprop = 0;
2015         icount = 0;
2016
2017         /*
2018          * Adjust the delete_tid.  Update the mirror_tid propagation field
2019          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2020          */
2021         if (delete_flags & HAMMER_DELETE_ADJUST) {
2022                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2023                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2024                                 icount = -1;
2025                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2026                                 icount = 1;
2027                 }
2028
2029                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2030                 elm->leaf.base.delete_tid = delete_tid;
2031                 elm->leaf.delete_ts = delete_ts;
2032                 hammer_modify_node_done(node);
2033
2034                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2035                         hammer_modify_node_field(trans, node, mirror_tid);
2036                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2037                         hammer_modify_node_done(node);
2038                         doprop = 1;
2039                         if (hammer_debug_general & 0x0002) {
2040                                 kprintf("delete_at_cursor: propagate %016llx"
2041                                         " @%016llx\n",
2042                                         elm->leaf.base.delete_tid,
2043                                         node->node_offset);
2044                         }
2045                 }
2046
2047                 /*
2048                  * Adjust for the iteration.  We have deleted the current
2049                  * element and want to clear ATEDISK so the iteration does
2050                  * not skip the element after, which now becomes the current
2051                  * element.
2052                  */
2053                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2054                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
2055                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2056                 }
2057
2058                 /*
2059                  * An on-disk record cannot have the same delete_tid
2060                  * as its create_tid.  In a chain of record updates
2061                  * this could result in a duplicate record.
2062                  */
2063                 KKASSERT(elm->leaf.base.delete_tid !=
2064                          elm->leaf.base.create_tid);
2065         }
2066
2067         /*
2068          * Destroy the B-Tree element if asked (typically if a nohistory
2069          * file or mount, or when called by the pruning code).
2070          *
2071          * Adjust the ATEDISK flag to properly support iterations.
2072          */
2073         if (delete_flags & HAMMER_DELETE_DESTROY) {
2074                 data_offset = elm->leaf.data_offset;
2075                 data_len = elm->leaf.data_len;
2076                 rec_type = elm->leaf.base.rec_type;
2077                 if (doprop) {
2078                         save_leaf = elm->leaf;
2079                         leaf = &save_leaf;
2080                 }
2081                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2082                     elm->leaf.base.delete_tid == 0) {
2083                         icount = -1;
2084                 }
2085
2086                 error = hammer_btree_delete(cursor);
2087                 if (error == 0) {
2088                         /*
2089                          * This forces a fixup for the iteration because
2090                          * the cursor is now either sitting at the 'next'
2091                          * element or sitting at the end of a leaf.
2092                          */
2093                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2094                                 cursor->flags |= HAMMER_CURSOR_DELBTREE;
2095                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2096                         }
2097                 }
2098                 if (error == 0) {
2099                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2100                         case HAMMER_ZONE_LARGE_DATA:
2101                         case HAMMER_ZONE_SMALL_DATA:
2102                         case HAMMER_ZONE_META:
2103                                 hammer_blockmap_free(trans,
2104                                                      data_offset, data_len);
2105                                 break;
2106                         default:
2107                                 break;
2108                         }
2109                 }
2110         }
2111
2112         /*
2113          * Track inode count and next_tid.  This is used by the mirroring
2114          * and PFS code.  icount can be negative, zero, or positive.
2115          */
2116         if (error == 0 && track) {
2117                 if (icount) {
2118                         hammer_modify_volume_field(trans, trans->rootvol,
2119                                                    vol0_stat_inodes);
2120                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2121                         hammer_modify_volume_done(trans->rootvol);
2122                 }
2123                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2124                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2125                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2126                         hammer_modify_volume_done(trans->rootvol);
2127                 }
2128         }
2129
2130         /*
2131          * mirror_tid propagation occurs if the node's mirror_tid had to be
2132          * updated while adjusting the delete_tid.
2133          *
2134          * This occurs when deleting even in nohistory mode, but does not
2135          * occur when pruning an already-deleted node.
2136          *
2137          * cursor->ip is NULL when called from the pruning, mirroring,
2138          * and pfs code.  If non-NULL propagation will be conditionalized
2139          * on whether the PFS is in no-history mode or not.
2140          */
2141         if (doprop) {
2142                 if (cursor->ip)
2143                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2144                 else
2145                         hammer_btree_do_propagation(cursor, NULL, leaf);
2146         }
2147         hammer_sync_unlock(trans);
2148         return (error);
2149 }
2150
2151 /*
2152  * Determine whether we can remove a directory.  This routine checks whether
2153  * a directory is empty or not and enforces flush connectivity.
2154  *
2155  * Flush connectivity requires that we block if the target directory is
2156  * currently flushing, otherwise it may not end up in the same flush group.
2157  *
2158  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2159  */
2160 int
2161 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2162 {
2163         struct hammer_cursor cursor;
2164         int error;
2165
2166         /*
2167          * Check directory empty
2168          */
2169         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2170
2171         cursor.key_beg.localization = ip->obj_localization +
2172                                       HAMMER_LOCALIZE_MISC;
2173         cursor.key_beg.obj_id = ip->obj_id;
2174         cursor.key_beg.create_tid = 0;
2175         cursor.key_beg.delete_tid = 0;
2176         cursor.key_beg.obj_type = 0;
2177         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2178         cursor.key_beg.key = HAMMER_MIN_KEY;
2179
2180         cursor.key_end = cursor.key_beg;
2181         cursor.key_end.rec_type = 0xFFFF;
2182         cursor.key_end.key = HAMMER_MAX_KEY;
2183
2184         cursor.asof = ip->obj_asof;
2185         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2186
2187         error = hammer_ip_first(&cursor);
2188         if (error == ENOENT)
2189                 error = 0;
2190         else if (error == 0)
2191                 error = ENOTEMPTY;
2192         hammer_done_cursor(&cursor);
2193         return(error);
2194 }
2195