sys/vfs/hammer/hammer_object.c

   1 /*
   2  * Copyright (c) 2007 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 1.31 2008/02/23 20:55:23 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_add(hammer_transaction_t trans, hammer_record_t record);
  40 static int hammer_mem_lookup(hammer_cursor_t cursor, hammer_inode_t ip);
  41 static int hammer_mem_first(hammer_cursor_t cursor, hammer_inode_t ip);
  42
  43 /*
  44  * Red-black tree support.
  45  */
  46 static int
  47 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  48 {
  49         if (rec1->rec.base.base.rec_type < rec2->rec.base.base.rec_type)
  50                 return(-1);
  51         if (rec1->rec.base.base.rec_type > rec2->rec.base.base.rec_type)
  52                 return(1);
  53
  54         if (rec1->rec.base.base.key < rec2->rec.base.base.key)
  55                 return(-1);
  56         if (rec1->rec.base.base.key > rec2->rec.base.base.key)
  57                 return(1);
  58
  59         if (rec1->rec.base.base.create_tid == 0) {
  60                 if (rec2->rec.base.base.create_tid == 0)
  61                         return(0);
  62                 return(1);
  63         }
  64         if (rec2->rec.base.base.create_tid == 0)
  65                 return(-1);
  66
  67         if (rec1->rec.base.base.create_tid < rec2->rec.base.base.create_tid)
  68                 return(-1);
  69         if (rec1->rec.base.base.create_tid > rec2->rec.base.base.create_tid)
  70                 return(1);
  71         return(0);
  72 }
  73
  74 static int
  75 hammer_rec_compare(hammer_base_elm_t info, hammer_record_t rec)
  76 {
  77         if (info->rec_type < rec->rec.base.base.rec_type)
  78                 return(-3);
  79         if (info->rec_type > rec->rec.base.base.rec_type)
  80                 return(3);
  81
  82         if (info->key < rec->rec.base.base.key)
  83                 return(-2);
  84         if (info->key > rec->rec.base.base.key)
  85                 return(2);
  86
  87         if (info->create_tid == 0) {
  88                 if (rec->rec.base.base.create_tid == 0)
  89                         return(0);
  90                 return(1);
  91         }
  92         if (rec->rec.base.base.create_tid == 0)
  93                 return(-1);
  94         if (info->create_tid < rec->rec.base.base.create_tid)
  95                 return(-1);
  96         if (info->create_tid > rec->rec.base.base.create_tid)
  97                 return(1);
  98         return(0);
  99 }
 100
 101 /*
 102  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 103  * is reversed so the comparison result has to be negated.  key_beg and
 104  * key_end are both range-inclusive.
 105  *
 106  * The creation timestamp can cause hammer_rec_compare() to return -1 or +1.
 107  * These do not stop the scan.
 108  *
 109  * Localized deletions are not cached in-memory.
 110  */
 111 static
 112 int
 113 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 114 {
 115         hammer_cursor_t cursor = data;
 116         int r;
 117
 118         r = hammer_rec_compare(&cursor->key_beg, rec);
 119         if (r > 1)
 120                 return(-1);
 121         r = hammer_rec_compare(&cursor->key_end, rec);
 122         if (r < -1)
 123                 return(1);
 124         return(0);
 125 }
 126
 127 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 128 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 129                     hammer_rec_compare, hammer_base_elm_t);
 130
 131 /*
 132  * Allocate a record for the caller to finish filling in.  The record is
 133  * returned referenced.
 134  */
 135 hammer_record_t
 136 hammer_alloc_mem_record(hammer_inode_t ip)
 137 {
 138         hammer_record_t record;
 139
 140         ++hammer_count_records;
 141         record = kmalloc(sizeof(*record), M_HAMMER, M_WAITOK|M_ZERO);
 142         record->ip = ip;
 143         record->rec.base.base.btype = HAMMER_BTREE_TYPE_RECORD;
 144         hammer_ref(&record->lock);
 145         return (record);
 146 }
 147
 148 /*
 149  * Release a memory record.  Records marked for deletion are immediately
 150  * removed from the RB-Tree but otherwise left intact until the last ref
 151  * goes away.
 152  */
 153 void
 154 hammer_rel_mem_record(struct hammer_record *record)
 155 {
 156         hammer_unref(&record->lock);
 157
 158         if (record->flags & HAMMER_RECF_DELETED) {
 159                 if (record->flags & HAMMER_RECF_ONRBTREE) {
 160                         RB_REMOVE(hammer_rec_rb_tree, &record->ip->rec_tree,
 161                                   record);
 162                         record->flags &= ~HAMMER_RECF_ONRBTREE;
 163                 }
 164                 if (record->lock.refs == 0) {
 165                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 166                                 --hammer_count_record_datas;
 167                                 kfree(record->data, M_HAMMER);
 168                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 169                         }
 170                         record->data = NULL;
 171                         --hammer_count_records;
 172                         kfree(record, M_HAMMER);
 173                         return;
 174                 }
 175         }
 176
 177         /*
 178          * If someone wanted the record wake them up.
 179          */
 180         if (record->flags & HAMMER_RECF_WANTED) {
 181                 record->flags &= ~HAMMER_RECF_WANTED;
 182                 wakeup(record);
 183         }
 184 }
 185
 186 /*
 187  * Lookup an in-memory record given the key specified in the cursor.  Works
 188  * just like hammer_btree_lookup() but operates on an inode's in-memory
 189  * record list.
 190  *
 191  * The lookup must fail if the record is marked for deferred deletion.
 192  */
 193 static
 194 int
 195 hammer_mem_lookup(hammer_cursor_t cursor, hammer_inode_t ip)
 196 {
 197         int error;
 198
 199         if (cursor->iprec) {
 200                 hammer_rel_mem_record(cursor->iprec);
 201                 cursor->iprec = NULL;
 202         }
 203         if (cursor->ip) {
 204                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 205                                                   &cursor->ip->rec_tree);
 206         }
 207         cursor->ip = ip;
 208         hammer_rec_rb_tree_scan_info_link(&cursor->scan, &ip->rec_tree);
 209         cursor->scan.node = NULL;
 210         cursor->iprec = hammer_rec_rb_tree_RB_LOOKUP_INFO(
 211                                 &ip->rec_tree, &cursor->key_beg);
 212         if (cursor->iprec == NULL) {
 213                 error = ENOENT;
 214         } else {
 215                 hammer_ref(&cursor->iprec->lock);
 216                 error = 0;
 217         }
 218         return(error);
 219 }
 220
 221 /*
 222  * hammer_mem_first() - locate the first in-memory record matching the
 223  * cursor.
 224  *
 225  * The RB_SCAN function we use is designed as a callback.  We terminate it
 226  * (return -1) as soon as we get a match.
 227  */
 228 static
 229 int
 230 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 231 {
 232         hammer_cursor_t cursor = data;
 233
 234         /*
 235          * We terminate on success, so this should be NULL on entry.
 236          */
 237         KKASSERT(cursor->iprec == NULL);
 238
 239         /*
 240          * Skip if the record was marked deleted
 241          */
 242         if (rec->flags & HAMMER_RECF_DELETED)
 243                 return(0);
 244
 245         /*
 246          * Skip if not visible due to our as-of TID
 247          */
 248         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 249                 if (cursor->asof < rec->rec.base.base.create_tid)
 250                         return(0);
 251                 if (rec->rec.base.base.delete_tid &&
 252                     cursor->asof >= rec->rec.base.base.delete_tid) {
 253                         return(0);
 254                 }
 255         }
 256
 257         /*
 258          * Block if currently being synchronized to disk, otherwise we
 259          * may get a duplicate.  Wakeup the syncer if it's stuck on
 260          * the record.
 261          */
 262         hammer_ref(&rec->lock);
 263         ++rec->blocked;
 264         while (rec->flags & HAMMER_RECF_SYNCING) {
 265                 rec->flags |= HAMMER_RECF_WANTED;
 266                 tsleep(rec, 0, "hmrrc2", 0);
 267         }
 268         --rec->blocked;
 269
 270         /*
 271          * The record may have been deleted while we were blocked.
 272          */
 273         if (rec->flags & HAMMER_RECF_DELETED) {
 274                 hammer_rel_mem_record(cursor->iprec);
 275                 return(0);
 276         }
 277
 278         /*
 279          * Set the matching record and stop the scan.
 280          */
 281         cursor->iprec = rec;
 282         return(-1);
 283 }
 284
 285 static
 286 int
 287 hammer_mem_first(hammer_cursor_t cursor, hammer_inode_t ip)
 288 {
 289         if (cursor->iprec) {
 290                 hammer_rel_mem_record(cursor->iprec);
 291                 cursor->iprec = NULL;
 292         }
 293         if (cursor->ip) {
 294                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 295                                                   &cursor->ip->rec_tree);
 296         }
 297         cursor->ip = ip;
 298         hammer_rec_rb_tree_scan_info_link(&cursor->scan, &ip->rec_tree);
 299
 300         cursor->scan.node = NULL;
 301         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 302                                    hammer_rec_scan_callback, cursor);
 303
 304         /*
 305          * Adjust scan.node and keep it linked into the RB-tree so we can
 306          * hold the cursor through third party modifications of the RB-tree.
 307          */
 308         if (cursor->iprec) {
 309                 cursor->scan.node = hammer_rec_rb_tree_RB_NEXT(cursor->iprec);
 310                 return(0);
 311         }
 312         return(ENOENT);
 313 }
 314
 315 void
 316 hammer_mem_done(hammer_cursor_t cursor)
 317 {
 318         if (cursor->ip) {
 319                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 320                                                   &cursor->ip->rec_tree);
 321                 cursor->ip = NULL;
 322         }
 323         if (cursor->iprec) {
 324                 hammer_rel_mem_record(cursor->iprec);
 325                 cursor->iprec = NULL;
 326         }
 327 }
 328
 329 /************************************************************************
 330  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 331  ************************************************************************
 332  *
 333  * These functions manipulate in-memory records.  Such records typically
 334  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 335  */
 336
 337 /*
 338  * Add a directory entry (dip,ncp) which references inode (ip).
 339  *
 340  * Note that the low 32 bits of the namekey are set temporarily to create
 341  * a unique in-memory record, and may be modified a second time when the
 342  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 343  * all 0's when synching to disk, which is not handled here.
 344  */
 345 int
 346 hammer_ip_add_directory(struct hammer_transaction *trans,
 347                      struct hammer_inode *dip, struct namecache *ncp,
 348                      struct hammer_inode *ip)
 349 {
 350         hammer_record_t record;
 351         int error;
 352         int bytes;
 353
 354         record = hammer_alloc_mem_record(dip);
 355
 356         bytes = ncp->nc_nlen;   /* NOTE: terminating \0 is NOT included */
 357         if (++trans->hmp->namekey_iterator == 0)
 358                 ++trans->hmp->namekey_iterator;
 359
 360         record->rec.entry.base.base.obj_id = dip->obj_id;
 361         record->rec.entry.base.base.key =
 362                 hammer_directory_namekey(ncp->nc_name, bytes);
 363         record->rec.entry.base.base.key += trans->hmp->namekey_iterator;
 364         record->rec.entry.base.base.create_tid = trans->tid;
 365         record->rec.entry.base.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 366         record->rec.entry.base.base.obj_type = ip->ino_rec.base.base.obj_type;
 367         record->rec.entry.obj_id = ip->obj_id;
 368         record->data = (void *)ncp->nc_name;
 369         record->rec.entry.base.data_len = bytes;
 370         ++ip->ino_rec.ino_nlinks;
 371         hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 372         /* NOTE: copies record->data */
 373         error = hammer_mem_add(trans, record);
 374         return(error);
 375 }
 376
 377 /*
 378  * Delete the directory entry and update the inode link count.  The
 379  * cursor must be seeked to the directory entry record being deleted.
 380  *
 381  * NOTE: HAMMER_CURSOR_DELETE may not have been set.  XXX remove flag.
 382  *
 383  * This function can return EDEADLK requiring the caller to terminate
 384  * the cursor and retry.
 385  */
 386 int
 387 hammer_ip_del_directory(struct hammer_transaction *trans,
 388                      hammer_cursor_t cursor, struct hammer_inode *dip,
 389                      struct hammer_inode *ip)
 390 {
 391         int error;
 392
 393         error = hammer_ip_delete_record(cursor, trans->tid);
 394
 395         /*
 396          * One less link.  The file may still be open in the OS even after
 397          * all links have gone away so we only try to sync if the OS has
 398          * no references and nlinks falls to 0.
 399          *
 400          * We have to terminate the cursor before syncing the inode to
 401          * avoid deadlocking against ourselves.
 402          */
 403         if (error == 0) {
 404                 --ip->ino_rec.ino_nlinks;
 405                 hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 406                 if (ip->ino_rec.ino_nlinks == 0 &&
 407                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 408                         hammer_done_cursor(cursor);
 409                         hammer_sync_inode(ip, MNT_NOWAIT, 1);
 410                 }
 411
 412         }
 413         return(error);
 414 }
 415
 416 /*
 417  * Add a record to an inode.
 418  *
 419  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 420  * initialize the following additional fields:
 421  *
 422  * record->rec.entry.base.base.key
 423  * record->rec.entry.base.base.rec_type
 424  * record->rec.entry.base.base.data_len
 425  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 426  */
 427 int
 428 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 429 {
 430         hammer_inode_t ip = record->ip;
 431         int error;
 432
 433         record->rec.base.base.obj_id = ip->obj_id;
 434         record->rec.base.base.create_tid = trans->tid;
 435         record->rec.base.base.obj_type = ip->ino_rec.base.base.obj_type;
 436
 437         hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 438         /* NOTE: copies record->data */
 439         error = hammer_mem_add(trans, record);
 440         return(error);
 441 }
 442
 443 /*
 444  * Sync data from a buffer cache buffer (typically) to the filesystem.  This
 445  * is called via the strategy called from a cached data source.  This code
 446  * is responsible for actually writing a data record out to the disk.
 447  *
 448  * This can only occur non-historically (i.e. 'current' data only).
 449  */
 450 int
 451 hammer_ip_sync_data(hammer_transaction_t trans, hammer_inode_t ip,
 452                        int64_t offset, void *data, int bytes)
 453 {
 454         struct hammer_cursor cursor;
 455         hammer_record_ondisk_t rec;
 456         union hammer_btree_elm elm;
 457         hammer_off_t rec_offset;
 458         void *bdata;
 459         int error;
 460
 461         KKASSERT((offset & HAMMER_BUFMASK) == 0);
 462         KKASSERT((bytes & HAMMER_BUFMASK) == 0);
 463 retry:
 464         error = hammer_init_cursor_hmp(&cursor, &ip->cache[0], ip->hmp);
 465         if (error)
 466                 return(error);
 467         cursor.key_beg.obj_id = ip->obj_id;
 468         cursor.key_beg.key = offset + bytes;
 469         cursor.key_beg.create_tid = trans->tid;
 470         cursor.key_beg.delete_tid = 0;
 471         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
 472         cursor.asof = trans->tid;
 473         cursor.flags |= HAMMER_CURSOR_INSERT;
 474
 475         /*
 476          * Issue a lookup to position the cursor.
 477          */
 478         error = hammer_btree_lookup(&cursor);
 479         if (error == 0) {
 480                 kprintf("hammer_ip_sync_data: duplicate data at (%lld,%d)\n",
 481                         offset, bytes);
 482                 hammer_print_btree_elm(&cursor.node->ondisk->elms[cursor.index],
 483                                        HAMMER_BTREE_TYPE_LEAF, cursor.index);
 484                 error = EIO;
 485         }
 486         if (error != ENOENT)
 487                 goto done;
 488
 489         /*
 490          * Allocate record and data space.  HAMMER_RECTYPE_DATA records
 491          * can cross buffer boundaries so we may have to split our bcopy.
 492          */
 493         rec = hammer_alloc_record(ip->hmp, &rec_offset, HAMMER_RECTYPE_DATA,
 494                                   &cursor.record_buffer,
 495                                   bytes, &bdata,
 496                                   &cursor.data_buffer, &error);
 497         if (rec == NULL)
 498                 goto done;
 499         if (hammer_debug_general & 0x1000)
 500                 kprintf("OOB RECOR2 DATA REC %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, rec->base.data_len);
 501
 502         /*
 503          * Fill everything in and insert our B-Tree node.
 504          *
 505          * NOTE: hammer_alloc_record() has already marked the related
 506          * buffers as modified.  If we do it again we will generate
 507          * unnecessary undo elements.
 508          */
 509         rec->base.base.btype = HAMMER_BTREE_TYPE_RECORD;
 510         rec->base.base.obj_id = ip->obj_id;
 511         rec->base.base.key = offset + bytes;
 512         rec->base.base.create_tid = trans->tid;
 513         rec->base.base.delete_tid = 0;
 514         rec->base.base.rec_type = HAMMER_RECTYPE_DATA;
 515         rec->base.data_crc = crc32(data, bytes);
 516         KKASSERT(rec->base.data_len == bytes);
 517
 518         bcopy(data, bdata, bytes);
 519
 520         elm.leaf.base = rec->base.base;
 521         elm.leaf.rec_offset = rec_offset;
 522         elm.leaf.data_offset = rec->base.data_off;
 523         elm.leaf.data_len = bytes;
 524         elm.leaf.data_crc = rec->base.data_crc;
 525
 526         /*
 527          * Data records can wind up on-disk before the inode itself is
 528          * on-disk.  One must assume data records may be on-disk if either
 529          * HAMMER_INODE_DONDISK or HAMMER_INODE_ONDISK is set
 530          */
 531         ip->flags |= HAMMER_INODE_DONDISK;
 532
 533         error = hammer_btree_insert(&cursor, &elm);
 534         if (error == 0)
 535                 goto done;
 536
 537         hammer_blockmap_free(ip->hmp, rec_offset, HAMMER_RECORD_SIZE);
 538 done:
 539         hammer_done_cursor(&cursor);
 540         if (error == EDEADLK)
 541                 goto retry;
 542         return(error);
 543 }
 544
 545 /*
 546  * Sync an in-memory record to the disk.  This is typically called via fsync
 547  * from a cached record source.  This code is responsible for actually
 548  * writing a record out to the disk.
 549  */
 550 int
 551 hammer_ip_sync_record(hammer_record_t record)
 552 {
 553         struct hammer_cursor cursor;
 554         hammer_record_ondisk_t rec;
 555         hammer_mount_t hmp;
 556         union hammer_btree_elm elm;
 557         hammer_off_t rec_offset;
 558         void *bdata;
 559         int error;
 560
 561         hmp = record->ip->hmp;
 562 retry:
 563         /*
 564          * If the record has been deleted or is being synchronized, stop.
 565          * Interlock with the syncing flag.
 566          */
 567         if (record->flags & (HAMMER_RECF_DELETED | HAMMER_RECF_SYNCING))
 568                 return(0);
 569         record->flags |= HAMMER_RECF_SYNCING;
 570
 571         /*
 572          * If someone other then us is referencing the record and not
 573          * blocking waiting for us, we have to wait until they finish.
 574          *
 575          * It is possible the record got destroyed while we were blocked.
 576          */
 577         if (record->lock.refs > record->blocked + 1) {
 578                 while (record->lock.refs > record->blocked + 1) {
 579                         record->flags |= HAMMER_RECF_WANTED;
 580                         tsleep(record, 0, "hmrrc1", 0);
 581                 }
 582                 if (record->flags & HAMMER_RECF_DELETED)
 583                         return(0);
 584         }
 585
 586         /*
 587          * Get a cursor
 588          */
 589         error = hammer_init_cursor_hmp(&cursor, &record->ip->cache[0], hmp);
 590         if (error)
 591                 return(error);
 592         cursor.key_beg = record->rec.base.base;
 593         cursor.flags |= HAMMER_CURSOR_INSERT;
 594
 595         /*
 596          * Issue a lookup to position the cursor and locate the cluster.  The
 597          * target key should not exist.  If we are creating a directory entry
 598          * we may have to iterate the low 32 bits of the key to find an unused
 599          * key.
 600          */
 601         for (;;) {
 602                 error = hammer_btree_lookup(&cursor);
 603                 if (error)
 604                         break;
 605                 if (record->rec.base.base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
 606                         kprintf("hammer_ip_sync_record: duplicate rec "
 607                                 "at (%016llx)\n", record->rec.base.base.key);
 608                         Debugger("duplicate record1");
 609                         error = EIO;
 610                         break;
 611                 }
 612                 if (++hmp->namekey_iterator == 0)
 613                         ++hmp->namekey_iterator;
 614                 record->rec.base.base.key &= ~(0xFFFFFFFFLL);
 615                 record->rec.base.base.key |= hmp->namekey_iterator;
 616                 cursor.key_beg.key = record->rec.base.base.key;
 617         }
 618         if (error != ENOENT)
 619                 goto done;
 620
 621         /*
 622          * Mark the record as undergoing synchronization.  Our cursor is
 623          * holding a locked B-Tree node for the insertion which interlocks
 624          * anyone trying to access this record.
 625          *
 626          * XXX There is still a race present related to iterations.  An
 627          * iteration may process the record, a sync may occur, and then
 628          * later process the B-Tree element for the same record.
 629          *
 630          * We do not try to synchronize a deleted record.
 631          */
 632         if (record->flags & HAMMER_RECF_DELETED) {
 633                 error = 0;
 634                 goto done;
 635         }
 636
 637         /*
 638          * Allocate the record and data.  The result buffers will be
 639          * marked as being modified and further calls to
 640          * hammer_modify_buffer() will result in unneeded UNDO records.
 641          *
 642          * Support zero-fill records (data == NULL and data_len != 0)
 643          */
 644         if (record->data == NULL) {
 645                 rec = hammer_alloc_record(hmp, &rec_offset,
 646                                           record->rec.base.base.rec_type,
 647                                           &cursor.record_buffer,
 648                                           0, &bdata,
 649                                           NULL, &error);
 650                 if (hammer_debug_general & 0x1000)
 651                         kprintf("NULL RECORD DATA\n");
 652         } else if (record->flags & HAMMER_RECF_INBAND) {
 653                 rec = hammer_alloc_record(hmp, &rec_offset,
 654                                           record->rec.base.base.rec_type,
 655                                           &cursor.record_buffer,
 656                                           record->rec.base.data_len, &bdata,
 657                                           NULL, &error);
 658                 if (hammer_debug_general & 0x1000)
 659                         kprintf("INBAND RECORD DATA %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, record->rec.base.data_len);
 660         } else {
 661                 rec = hammer_alloc_record(hmp, &rec_offset,
 662                                           record->rec.base.base.rec_type,
 663                                           &cursor.record_buffer,
 664                                           record->rec.base.data_len, &bdata,
 665                                           &cursor.data_buffer, &error);
 666                 if (hammer_debug_general & 0x1000)
 667                         kprintf("OOB RECORD DATA REC %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, record->rec.base.data_len);
 668         }
 669
 670         if (rec == NULL)
 671                 goto done;
 672
 673         /*
 674          * Fill in the remaining fields and insert our B-Tree node.
 675          */
 676         rec->base.base = record->rec.base.base;
 677         bcopy(&record->rec.base + 1, &rec->base + 1,
 678               HAMMER_RECORD_SIZE - sizeof(record->rec.base));
 679
 680         /*
 681          * Copy the data and deal with zero-fill support.
 682          */
 683         if (record->data) {
 684                 rec->base.data_crc = crc32(record->data, rec->base.data_len);
 685                 bcopy(record->data, bdata, rec->base.data_len);
 686         } else {
 687                 rec->base.data_len = record->rec.base.data_len;
 688         }
 689
 690         elm.leaf.base = record->rec.base.base;
 691         elm.leaf.rec_offset = rec_offset;
 692         elm.leaf.data_offset = rec->base.data_off;
 693         elm.leaf.data_len = rec->base.data_len;
 694         elm.leaf.data_crc = rec->base.data_crc;
 695
 696         error = hammer_btree_insert(&cursor, &elm);
 697
 698         /*
 699          * Clean up on success, or fall through on error.
 700          */
 701         if (error == 0) {
 702                 record->flags |= HAMMER_RECF_DELETED;
 703                 goto done;
 704         }
 705
 706         /*
 707          * Try to unwind the allocation
 708          */
 709         hammer_blockmap_free(hmp, rec_offset, HAMMER_RECORD_SIZE);
 710 done:
 711         record->flags &= ~HAMMER_RECF_SYNCING;
 712         hammer_done_cursor(&cursor);
 713         if (error == EDEADLK)
 714                 goto retry;
 715         return(error);
 716 }
 717
 718 /*
 719  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
 720  * entry's key is used to deal with hash collisions in the upper 32 bits.
 721  * A unique 64 bit key is generated in-memory and may be regenerated a
 722  * second time when the directory record is flushed to the on-disk B-Tree.
 723  *
 724  * A referenced record is passed to this function.  This function
 725  * eats the reference.  If an error occurs the record will be deleted.
 726  *
 727  * A copy of the temporary record->data pointer provided by the caller
 728  * will be made.
 729  */
 730 static
 731 int
 732 hammer_mem_add(struct hammer_transaction *trans, hammer_record_t record)
 733 {
 734         void *data;
 735         int bytes;
 736         int reclen;
 737
 738         /*
 739          * Make a private copy of record->data
 740          */
 741         if (record->data) {
 742                 /*
 743                  * Try to embed the data in extra space in the record
 744                  * union, otherwise allocate a copy.
 745                  */
 746                 bytes = record->rec.base.data_len;
 747                 switch(record->rec.base.base.rec_type) {
 748                 case HAMMER_RECTYPE_DIRENTRY:
 749                         reclen = offsetof(struct hammer_entry_record, name[0]);
 750                         break;
 751                 case HAMMER_RECTYPE_DATA:
 752                         reclen = offsetof(struct hammer_data_record, data[0]);
 753                         break;
 754                 default:
 755                         reclen = sizeof(record->rec);
 756                         break;
 757                 }
 758                 if (reclen + bytes <= HAMMER_RECORD_SIZE) {
 759                         bcopy(record->data, (char *)&record->rec + reclen,
 760                               bytes);
 761                         record->data = (void *)((char *)&record->rec + reclen);
 762                         record->flags |= HAMMER_RECF_INBAND;
 763                 } else {
 764                         ++hammer_count_record_datas;
 765                         data = kmalloc(bytes, M_HAMMER, M_WAITOK);
 766                         record->flags |= HAMMER_RECF_ALLOCDATA;
 767                         bcopy(record->data, data, bytes);
 768                         record->data = data;
 769                 }
 770         }
 771
 772         /*
 773          * Insert into the RB tree, find an unused iterator if this is
 774          * a directory entry.
 775          */
 776         while (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
 777                 if (record->rec.base.base.rec_type != HAMMER_RECTYPE_DIRENTRY){
 778                         record->flags |= HAMMER_RECF_DELETED;
 779                         hammer_rel_mem_record(record);
 780                         return (EEXIST);
 781                 }
 782                 if (++trans->hmp->namekey_iterator == 0)
 783                         ++trans->hmp->namekey_iterator;
 784                 record->rec.base.base.key &= ~(0xFFFFFFFFLL);
 785                 record->rec.base.base.key |= trans->hmp->namekey_iterator;
 786         }
 787         record->flags |= HAMMER_RECF_ONRBTREE;
 788         hammer_modify_inode(trans, record->ip, HAMMER_INODE_XDIRTY);
 789         hammer_rel_mem_record(record);
 790         return(0);
 791 }
 792
 793 /************************************************************************
 794  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
 795  ************************************************************************
 796  *
 797  * These functions augment the B-Tree scanning functions in hammer_btree.c
 798  * by merging in-memory records with on-disk records.
 799  */
 800
 801 /*
 802  * Locate a particular record either in-memory or on-disk.
 803  *
 804  * NOTE: This is basically a standalone routine, hammer_ip_next() may
 805  * NOT be called to iterate results.
 806  */
 807 int
 808 hammer_ip_lookup(hammer_cursor_t cursor, struct hammer_inode *ip)
 809 {
 810         int error;
 811
 812         /*
 813          * If the element is in-memory return it without searching the
 814          * on-disk B-Tree
 815          */
 816         error = hammer_mem_lookup(cursor, ip);
 817         if (error == 0) {
 818                 cursor->record = &cursor->iprec->rec;
 819                 return(error);
 820         }
 821         if (error != ENOENT)
 822                 return(error);
 823
 824         /*
 825          * If the inode has on-disk components search the on-disk B-Tree.
 826          */
 827         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
 828                 return(error);
 829         error = hammer_btree_lookup(cursor);
 830         if (error == 0)
 831                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
 832         return(error);
 833 }
 834
 835 /*
 836  * Locate the first record within the cursor's key_beg/key_end range,
 837  * restricted to a particular inode.  0 is returned on success, ENOENT
 838  * if no records matched the requested range, or some other error.
 839  *
 840  * When 0 is returned hammer_ip_next() may be used to iterate additional
 841  * records within the requested range.
 842  *
 843  * This function can return EDEADLK, requiring the caller to terminate
 844  * the cursor and try again.
 845  */
 846 int
 847 hammer_ip_first(hammer_cursor_t cursor, struct hammer_inode *ip)
 848 {
 849         int error;
 850
 851         /*
 852          * Clean up fields and setup for merged scan
 853          */
 854         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
 855         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
 856         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
 857         if (cursor->iprec) {
 858                 hammer_rel_mem_record(cursor->iprec);
 859                 cursor->iprec = NULL;
 860         }
 861
 862         /*
 863          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
 864          * exact lookup so if we get ENOENT we have to call the iterate
 865          * function to validate the first record after the begin key.
 866          *
 867          * The ATEDISK flag is used by hammer_btree_iterate to determine
 868          * whether it must index forwards or not.  It is also used here
 869          * to select the next record from in-memory or on-disk.
 870          *
 871          * EDEADLK can only occur if the lookup hit an empty internal
 872          * element and couldn't delete it.  Since this could only occur
 873          * in-range, we can just iterate from the failure point.
 874          */
 875         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
 876                 error = hammer_btree_lookup(cursor);
 877                 if (error == ENOENT || error == EDEADLK) {
 878                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 879                         error = hammer_btree_iterate(cursor);
 880                 }
 881                 if (error && error != ENOENT)
 882                         return(error);
 883                 if (error == 0) {
 884                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
 885                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 886                 } else {
 887                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 888                 }
 889         }
 890
 891         /*
 892          * Search the in-memory record list (Red-Black tree).  Unlike the
 893          * B-Tree search, mem_first checks for records in the range.
 894          */
 895         error = hammer_mem_first(cursor, ip);
 896         if (error && error != ENOENT)
 897                 return(error);
 898         if (error == 0) {
 899                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
 900                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
 901         }
 902
 903         /*
 904          * This will return the first matching record.
 905          */
 906         return(hammer_ip_next(cursor));
 907 }
 908
 909 /*
 910  * Retrieve the next record in a merged iteration within the bounds of the
 911  * cursor.  This call may be made multiple times after the cursor has been
 912  * initially searched with hammer_ip_first().
 913  *
 914  * 0 is returned on success, ENOENT if no further records match the
 915  * requested range, or some other error code is returned.
 916  */
 917 int
 918 hammer_ip_next(hammer_cursor_t cursor)
 919 {
 920         hammer_btree_elm_t elm;
 921         hammer_record_t rec;
 922         int error;
 923         int r;
 924
 925         /*
 926          * Load the current on-disk and in-memory record.  If we ate any
 927          * records we have to get the next one.
 928          *
 929          * If we deleted the last on-disk record we had scanned ATEDISK will
 930          * be clear and DELBTREE will be set, forcing a call to iterate. The
 931          * fact that ATEDISK is clear causes iterate to re-test the 'current'
 932          * element.  If ATEDISK is set, iterate will skip the 'current'
 933          * element.
 934          *
 935          * Get the next on-disk record
 936          */
 937         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
 938                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
 939                         error = hammer_btree_iterate(cursor);
 940                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
 941                         if (error == 0)
 942                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
 943                         else
 944                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
 945                                                  HAMMER_CURSOR_ATEDISK;
 946                 }
 947         }
 948
 949         /*
 950          * Get the next in-memory record.  The record can be ripped out
 951          * of the RB tree so we maintain a scan_info structure to track
 952          * the next node.
 953          *
 954          * hammer_rec_scan_cmp:  Is the record still in our general range,
 955          *                       (non-inclusive of snapshot exclusions)?
 956          * hammer_rec_scan_callback: Is the record in our snapshot?
 957          */
 958         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
 959                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
 960                         if (cursor->iprec) {
 961                                 hammer_rel_mem_record(cursor->iprec);
 962                                 cursor->iprec = NULL;
 963                         }
 964                         rec = cursor->scan.node;        /* next node */
 965                         while (rec) {
 966                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
 967                                         break;
 968                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
 969                                         break;
 970                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
 971                         }
 972                         if (cursor->iprec) {
 973                                 KKASSERT(cursor->iprec == rec);
 974                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
 975                                 cursor->scan.node =
 976                                         hammer_rec_rb_tree_RB_NEXT(rec);
 977                         } else {
 978                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
 979                         }
 980                 }
 981         }
 982
 983         /*
 984          * Extract either the disk or memory record depending on their
 985          * relative position.
 986          */
 987         error = 0;
 988         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
 989         case 0:
 990                 /*
 991                  * Both entries valid
 992                  */
 993                 elm = &cursor->node->ondisk->elms[cursor->index];
 994                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->rec.base.base);
 995                 if (r < 0) {
 996                         error = hammer_btree_extract(cursor,
 997                                                      HAMMER_CURSOR_GET_RECORD);
 998                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
 999                         break;
1000                 }
1001                 /* fall through to the memory entry */
1002         case HAMMER_CURSOR_ATEDISK:
1003                 /*
1004                  * Only the memory entry is valid
1005                  */
1006                 cursor->record = &cursor->iprec->rec;
1007                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1008                 break;
1009         case HAMMER_CURSOR_ATEMEM:
1010                 /*
1011                  * Only the disk entry is valid
1012                  */
1013                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
1014                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1015                 break;
1016         default:
1017                 /*
1018                  * Neither entry is valid
1019                  *
1020                  * XXX error not set properly
1021                  */
1022                 cursor->record = NULL;
1023                 error = ENOENT;
1024                 break;
1025         }
1026         return(error);
1027 }
1028
1029 /*
1030  * Resolve the cursor->data pointer for the current cursor position in
1031  * a merged iteration.
1032  */
1033 int
1034 hammer_ip_resolve_data(hammer_cursor_t cursor)
1035 {
1036         int error;
1037
1038         if (cursor->iprec && cursor->record == &cursor->iprec->rec) {
1039                 cursor->data = cursor->iprec->data;
1040                 error = 0;
1041         } else {
1042                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1043         }
1044         return(error);
1045 }
1046
1047 int
1048 hammer_ip_resolve_record_and_data(hammer_cursor_t cursor)
1049 {
1050         int error;
1051
1052         if (cursor->iprec && cursor->record == &cursor->iprec->rec) {
1053                 cursor->data = cursor->iprec->data;
1054                 error = 0;
1055         } else {
1056                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
1057                                                      HAMMER_CURSOR_GET_RECORD);
1058         }
1059         return(error);
1060 }
1061
1062 /*
1063  * Delete all records within the specified range for inode ip.
1064  *
1065  * NOTE: An unaligned range will cause new records to be added to cover
1066  * the edge cases. (XXX not implemented yet).
1067  *
1068  * NOTE: ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1069  *
1070  * NOTE: Record keys for regular file data have to be special-cased since
1071  * they indicate the end of the range (key = base + bytes).
1072  */
1073 int
1074 hammer_ip_delete_range(hammer_transaction_t trans, hammer_inode_t ip,
1075                        int64_t ran_beg, int64_t ran_end)
1076 {
1077         struct hammer_cursor cursor;
1078         hammer_record_ondisk_t rec;
1079         hammer_base_elm_t base;
1080         int error;
1081         int64_t off;
1082
1083 retry:
1084         hammer_init_cursor_hmp(&cursor, &ip->cache[0], ip->hmp);
1085
1086         cursor.key_beg.obj_id = ip->obj_id;
1087         cursor.key_beg.create_tid = 0;
1088         cursor.key_beg.delete_tid = 0;
1089         cursor.key_beg.obj_type = 0;
1090         cursor.asof = ip->obj_asof;
1091         cursor.flags |= HAMMER_CURSOR_ASOF;
1092
1093         cursor.key_end = cursor.key_beg;
1094         if (ip->ino_rec.base.base.obj_type == HAMMER_OBJTYPE_DBFILE) {
1095                 cursor.key_beg.key = ran_beg;
1096                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
1097                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
1098                 cursor.key_end.key = ran_end;
1099         } else {
1100                 /*
1101                  * The key in the B-Tree is (base+bytes), so the first possible
1102                  * matching key is ran_beg + 1.
1103                  */
1104                 int64_t tmp64;
1105
1106                 cursor.key_beg.key = ran_beg + 1;
1107                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
1108                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
1109
1110                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1111                 if (tmp64 < ran_end)
1112                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1113                 else
1114                         cursor.key_end.key = ran_end + MAXPHYS + 1;
1115         }
1116         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
1117
1118         error = hammer_ip_first(&cursor, ip);
1119
1120         /*
1121          * Iterate through matching records and mark them as deleted.
1122          */
1123         while (error == 0) {
1124                 rec = cursor.record;
1125                 base = &rec->base.base;
1126
1127                 KKASSERT(base->delete_tid == 0);
1128
1129                 /*
1130                  * There may be overlap cases for regular file data.  Also
1131                  * remember the key for a regular file record is the offset
1132                  * of the last byte of the record (base + len - 1), NOT the
1133                  * base offset.
1134                  */
1135 #if 0
1136                 kprintf("delete_range rec_type %02x\n", base->rec_type);
1137 #endif
1138                 if (base->rec_type == HAMMER_RECTYPE_DATA) {
1139 #if 0
1140                         kprintf("delete_range loop key %016llx\n",
1141                                 base->key - rec->base.data_len);
1142 #endif
1143                         off = base->key - rec->base.data_len;
1144                         /*
1145                          * Check the left edge case.  We currently do not
1146                          * split existing records.
1147                          */
1148                         if (off < ran_beg) {
1149                                 panic("hammer left edge case %016llx %d\n",
1150                                         base->key, rec->base.data_len);
1151                         }
1152
1153                         /*
1154                          * Check the right edge case.  Note that the
1155                          * record can be completely out of bounds, which
1156                          * terminates the search.
1157                          *
1158                          * base->key is exclusive of the right edge while
1159                          * ran_end is inclusive of the right edge.  The
1160                          * (key - data_len) left boundary is inclusive.
1161                          *
1162                          * XXX theory-check this test at some point, are
1163                          * we missing a + 1 somewhere?  Note that ran_end
1164                          * could overflow.
1165                          */
1166                         if (base->key - 1 > ran_end) {
1167                                 if (base->key - rec->base.data_len > ran_end)
1168                                         break;
1169                                 panic("hammer right edge case\n");
1170                         }
1171                 }
1172
1173                 /*
1174                  * Mark the record and B-Tree entry as deleted.  This will
1175                  * also physically delete the B-Tree entry, record, and
1176                  * data if the retention policy dictates.  The function
1177                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1178                  * uses to perform a fixup.
1179                  */
1180                 error = hammer_ip_delete_record(&cursor, trans->tid);
1181                 if (error)
1182                         break;
1183                 error = hammer_ip_next(&cursor);
1184         }
1185         hammer_done_cursor(&cursor);
1186         if (error == EDEADLK)
1187                 goto retry;
1188         if (error == ENOENT)
1189                 error = 0;
1190         return(error);
1191 }
1192
1193 /*
1194  * Delete all records associated with an inode except the inode record
1195  * itself.
1196  */
1197 int
1198 hammer_ip_delete_range_all(hammer_transaction_t trans, hammer_inode_t ip)
1199 {
1200         struct hammer_cursor cursor;
1201         hammer_record_ondisk_t rec;
1202         hammer_base_elm_t base;
1203         int error;
1204
1205 retry:
1206         hammer_init_cursor_hmp(&cursor, &ip->cache[0], ip->hmp);
1207
1208         cursor.key_beg.obj_id = ip->obj_id;
1209         cursor.key_beg.create_tid = 0;
1210         cursor.key_beg.delete_tid = 0;
1211         cursor.key_beg.obj_type = 0;
1212         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1213         cursor.key_beg.key = HAMMER_MIN_KEY;
1214
1215         cursor.key_end = cursor.key_beg;
1216         cursor.key_end.rec_type = 0xFFFF;
1217         cursor.key_end.key = HAMMER_MAX_KEY;
1218
1219         cursor.asof = ip->obj_asof;
1220         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1221
1222         error = hammer_ip_first(&cursor, ip);
1223
1224         /*
1225          * Iterate through matching records and mark them as deleted.
1226          */
1227         while (error == 0) {
1228                 rec = cursor.record;
1229                 base = &rec->base.base;
1230
1231                 KKASSERT(base->delete_tid == 0);
1232
1233                 /*
1234                  * Mark the record and B-Tree entry as deleted.  This will
1235                  * also physically delete the B-Tree entry, record, and
1236                  * data if the retention policy dictates.  The function
1237                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1238                  * uses to perform a fixup.
1239                  */
1240                 error = hammer_ip_delete_record(&cursor, trans->tid);
1241                 if (error)
1242                         break;
1243                 error = hammer_ip_next(&cursor);
1244         }
1245         hammer_done_cursor(&cursor);
1246         if (error == EDEADLK)
1247                 goto retry;
1248         if (error == ENOENT)
1249                 error = 0;
1250         return(error);
1251 }
1252
1253 /*
1254  * Delete the record at the current cursor.  On success the cursor will
1255  * be positioned appropriately for an iteration but may no longer be at
1256  * a leaf node.
1257  *
1258  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1259  * cursor and retry.
1260  */
1261 int
1262 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_tid_t tid)
1263 {
1264         hammer_btree_elm_t elm;
1265         hammer_mount_t hmp;
1266         int error;
1267         int dodelete;
1268
1269         /*
1270          * In-memory (unsynchronized) records can simply be freed.
1271          */
1272         if (cursor->record == &cursor->iprec->rec) {
1273                 cursor->iprec->flags |= HAMMER_RECF_DELETED;
1274                 return(0);
1275         }
1276
1277         /*
1278          * On-disk records are marked as deleted by updating their delete_tid.
1279          * This does not effect their position in the B-Tree (which is based
1280          * on their create_tid).
1281          */
1282         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
1283         elm = NULL;
1284         hmp = cursor->node->hmp;
1285
1286         dodelete = 0;
1287         if (error == 0) {
1288                 error = hammer_cursor_upgrade(cursor);
1289                 if (error == 0) {
1290                         hammer_modify_node(cursor->node);
1291                         elm = &cursor->node->ondisk->elms[cursor->index];
1292                         elm->leaf.base.delete_tid = tid;
1293                         hammer_modify_buffer(cursor->record_buffer, &cursor->record->base.base.delete_tid, sizeof(hammer_tid_t));
1294                         cursor->record->base.base.delete_tid = tid;
1295                 }
1296         }
1297
1298         /*
1299          * If we were mounted with the nohistory option, we physically
1300          * delete the record.
1301          */
1302         if (hmp->hflags & HMNT_NOHISTORY)
1303                 dodelete = 1;
1304
1305         if (error == 0 && dodelete) {
1306                 error = hammer_delete_at_cursor(cursor, NULL);
1307                 if (error) {
1308                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1309                         error = 0;
1310                 }
1311         }
1312         return(error);
1313 }
1314
1315 int
1316 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
1317 {
1318         hammer_btree_elm_t elm;
1319         hammer_off_t rec_offset;
1320         hammer_off_t data_offset;
1321         int32_t data_len;
1322         u_int8_t rec_type;
1323         int error;
1324
1325         elm = &cursor->node->ondisk->elms[cursor->index];
1326         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1327
1328         rec_offset = elm->leaf.rec_offset;
1329         data_offset = elm->leaf.data_offset;
1330         data_len = elm->leaf.data_len;
1331         rec_type = elm->leaf.base.rec_type;
1332
1333         error = hammer_btree_delete(cursor);
1334         if (error == 0) {
1335                 /*
1336                  * This forces a fixup for the iteration because
1337                  * the cursor is now either sitting at the 'next'
1338                  * element or sitting at the end of a leaf.
1339                  */
1340                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1341                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
1342                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1343                 }
1344         }
1345         if (error == 0) {
1346                 hammer_blockmap_free(cursor->node->hmp, rec_offset,
1347                                      sizeof(union hammer_record_ondisk));
1348         }
1349         if (error == 0) {
1350                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
1351                 case HAMMER_ZONE_LARGE_DATA:
1352                 case HAMMER_ZONE_SMALL_DATA:
1353                         hammer_blockmap_free(cursor->node->hmp,
1354                                              data_offset, data_len);
1355                         break;
1356                 default:
1357                         break;
1358                 }
1359         }
1360 #if 0
1361         kprintf("hammer_delete_at_cursor: %d:%d:%08x %08x/%d "
1362                 "(%d remain in cluster)\n",
1363                 cluster->volume->vol_no, cluster->clu_no,
1364                 rec_offset, data_offset, data_len,
1365                 cluster->ondisk->stat_records);
1366 #endif
1367         return (error);
1368 }
1369
1370 /*
1371  * Determine whether a directory is empty or not.  Returns 0 if the directory
1372  * is empty, ENOTEMPTY if it isn't, plus other possible errors.
1373  */
1374 int
1375 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
1376 {
1377         struct hammer_cursor cursor;
1378         int error;
1379
1380         hammer_init_cursor_hmp(&cursor, &ip->cache[0], ip->hmp);
1381
1382         cursor.key_beg.obj_id = ip->obj_id;
1383         cursor.key_beg.create_tid = 0;
1384         cursor.key_beg.delete_tid = 0;
1385         cursor.key_beg.obj_type = 0;
1386         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1387         cursor.key_beg.key = HAMMER_MIN_KEY;
1388
1389         cursor.key_end = cursor.key_beg;
1390         cursor.key_end.rec_type = 0xFFFF;
1391         cursor.key_end.key = HAMMER_MAX_KEY;
1392
1393         cursor.asof = ip->obj_asof;
1394         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1395
1396         error = hammer_ip_first(&cursor, ip);
1397         if (error == ENOENT)
1398                 error = 0;
1399         else if (error == 0)
1400                 error = ENOTEMPTY;
1401         hammer_done_cursor(&cursor);
1402         return(error);
1403 }
1404