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 1.75 2008/06/24 17:38:17 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_rec_trunc_callback(hammer_record_t record,
  43                                 void *data __unused);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45
  46 struct rec_trunc_info {
  47         u_int16_t       rec_type;
  48         int64_t         trunc_off;
  49 };
  50
  51 /*
  52  * Red-black tree support.  Comparison code for insertion.
  53  */
  54 static int
  55 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  56 {
  57         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  58                 return(-1);
  59         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  60                 return(1);
  61
  62         if (rec1->leaf.base.key < rec2->leaf.base.key)
  63                 return(-1);
  64         if (rec1->leaf.base.key > rec2->leaf.base.key)
  65                 return(1);
  66
  67         /*
  68          * Never match against an item deleted by the front-end.
  69          *
  70          * rec1 is greater then rec2 if rec1 is marked deleted.
  71          * rec1 is less then rec2 if rec2 is marked deleted.
  72          *
  73          * Multiple deleted records may be present, do not return 0
  74          * if both are marked deleted.
  75          */
  76         if (rec1->flags & HAMMER_RECF_DELETED_FE)
  77                 return(1);
  78         if (rec2->flags & HAMMER_RECF_DELETED_FE)
  79                 return(-1);
  80
  81         return(0);
  82 }
  83
  84 /*
  85  * Basic record comparison code similar to hammer_btree_cmp().
  86  */
  87 static int
  88 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
  89 {
  90         if (elm->rec_type < rec->leaf.base.rec_type)
  91                 return(-3);
  92         if (elm->rec_type > rec->leaf.base.rec_type)
  93                 return(3);
  94
  95         if (elm->key < rec->leaf.base.key)
  96                 return(-2);
  97         if (elm->key > rec->leaf.base.key)
  98                 return(2);
  99
 100         /*
 101          * Never match against an item deleted by the front-end.
 102          * elm is less then rec if rec is marked deleted.
 103          */
 104         if (rec->flags & HAMMER_RECF_DELETED_FE)
 105                 return(-1);
 106         return(0);
 107 }
 108
 109 /*
 110  * Special LOOKUP_INFO to locate an overlapping record.  This used by
 111  * the reservation code to implement small-block records (whos keys will
 112  * be different depending on data_len, when representing the same base
 113  * offset).
 114  *
 115  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 116  */
 117 static int
 118 hammer_rec_overlap_compare(hammer_btree_leaf_elm_t leaf, hammer_record_t rec)
 119 {
 120         if (leaf->base.rec_type < rec->leaf.base.rec_type)
 121                 return(-3);
 122         if (leaf->base.rec_type > rec->leaf.base.rec_type)
 123                 return(3);
 124
 125         /*
 126          * Overlap compare
 127          */
 128         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 129                 /* leaf_end <= rec_beg */
 130                 if (leaf->base.key <= rec->leaf.base.key - rec->leaf.data_len)
 131                         return(-2);
 132                 /* leaf_beg >= rec_end */
 133                 if (leaf->base.key - leaf->data_len >= rec->leaf.base.key)
 134                         return(2);
 135         } else {
 136                 if (leaf->base.key < rec->leaf.base.key)
 137                         return(-2);
 138                 if (leaf->base.key > rec->leaf.base.key)
 139                         return(2);
 140         }
 141
 142         /*
 143          * Never match against an item deleted by the front-end.
 144          * leaf is less then rec if rec is marked deleted.
 145          *
 146          * We must still return the proper code for the scan to continue
 147          * along the correct branches.
 148          */
 149         if (rec->flags & HAMMER_RECF_DELETED_FE) {
 150                 if (leaf->base.key < rec->leaf.base.key)
 151                         return(-2);
 152                 if (leaf->base.key > rec->leaf.base.key)
 153                         return(2);
 154                 return(-1);
 155         }
 156         return(0);
 157 }
 158
 159 /*
 160  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 161  * is reversed so the comparison result has to be negated.  key_beg and
 162  * key_end are both range-inclusive.
 163  *
 164  * Localized deletions are not cached in-memory.
 165  */
 166 static
 167 int
 168 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 169 {
 170         hammer_cursor_t cursor = data;
 171         int r;
 172
 173         r = hammer_rec_cmp(&cursor->key_beg, rec);
 174         if (r > 1)
 175                 return(-1);
 176         r = hammer_rec_cmp(&cursor->key_end, rec);
 177         if (r < -1)
 178                 return(1);
 179         return(0);
 180 }
 181
 182 /*
 183  * This compare function is used when simply looking up key_beg.
 184  */
 185 static
 186 int
 187 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 188 {
 189         hammer_cursor_t cursor = data;
 190         int r;
 191
 192         r = hammer_rec_cmp(&cursor->key_beg, rec);
 193         if (r > 1)
 194                 return(-1);
 195         if (r < -1)
 196                 return(1);
 197         return(0);
 198 }
 199
 200 /*
 201  * Locate blocks within the truncation range.  Partial blocks do not count.
 202  */
 203 static
 204 int
 205 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 206 {
 207         struct rec_trunc_info *info = data;
 208
 209         if (rec->leaf.base.rec_type < info->rec_type)
 210                 return(-1);
 211         if (rec->leaf.base.rec_type > info->rec_type)
 212                 return(1);
 213
 214         switch(rec->leaf.base.rec_type) {
 215         case HAMMER_RECTYPE_DB:
 216                 /*
 217                  * DB record key is not beyond the truncation point, retain.
 218                  */
 219                 if (rec->leaf.base.key < info->trunc_off)
 220                         return(-1);
 221                 break;
 222         case HAMMER_RECTYPE_DATA:
 223                 /*
 224                  * DATA record offset start is not beyond the truncation point,
 225                  * retain.
 226                  */
 227                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 228                         return(-1);
 229                 break;
 230         default:
 231                 panic("hammer_rec_trunc_cmp: unexpected record type");
 232         }
 233
 234         /*
 235          * The record start is >= the truncation point, return match,
 236          * the record should be destroyed.
 237          */
 238         return(0);
 239 }
 240
 241 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 242 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 243                     hammer_rec_overlap_compare, hammer_btree_leaf_elm_t);
 244
 245 /*
 246  * Allocate a record for the caller to finish filling in.  The record is
 247  * returned referenced.
 248  */
 249 hammer_record_t
 250 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 251 {
 252         hammer_record_t record;
 253
 254         ++hammer_count_records;
 255         record = kmalloc(sizeof(*record), M_HAMMER, M_WAITOK | M_ZERO);
 256         record->flush_state = HAMMER_FST_IDLE;
 257         record->ip = ip;
 258         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 259         record->leaf.data_len = data_len;
 260         hammer_ref(&record->lock);
 261
 262         if (data_len) {
 263                 record->data = kmalloc(data_len, M_HAMMER, M_WAITOK | M_ZERO);
 264                 record->flags |= HAMMER_RECF_ALLOCDATA;
 265                 ++hammer_count_record_datas;
 266         }
 267
 268         return (record);
 269 }
 270
 271 void
 272 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 273 {
 274         while (record->flush_state == HAMMER_FST_FLUSH) {
 275                 record->flags |= HAMMER_RECF_WANTED;
 276                 tsleep(record, 0, ident, 0);
 277         }
 278 }
 279
 280 /*
 281  * Called from the backend, hammer_inode.c, after a record has been
 282  * flushed to disk.  The record has been exclusively locked by the
 283  * caller and interlocked with BE.
 284  *
 285  * We clean up the state, unlock, and release the record (the record
 286  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 287  */
 288 void
 289 hammer_flush_record_done(hammer_record_t record, int error)
 290 {
 291         hammer_inode_t target_ip;
 292
 293         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 294         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 295
 296         if (error) {
 297                 /*
 298                  * An error occured, the backend was unable to sync the
 299                  * record to its media.  Leave the record intact.
 300                  */
 301                 Debugger("flush_record_done error");
 302         }
 303
 304         if (record->flags & HAMMER_RECF_DELETED_BE) {
 305                 if ((target_ip = record->target_ip) != NULL) {
 306                         TAILQ_REMOVE(&target_ip->target_list, record,
 307                                      target_entry);
 308                         record->target_ip = NULL;
 309                         hammer_test_inode(target_ip);
 310                 }
 311                 record->flush_state = HAMMER_FST_IDLE;
 312         } else {
 313                 if (record->target_ip) {
 314                         record->flush_state = HAMMER_FST_SETUP;
 315                         hammer_test_inode(record->ip);
 316                         hammer_test_inode(record->target_ip);
 317                 } else {
 318                         record->flush_state = HAMMER_FST_IDLE;
 319                 }
 320         }
 321         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 322         if (record->flags & HAMMER_RECF_WANTED) {
 323                 record->flags &= ~HAMMER_RECF_WANTED;
 324                 wakeup(record);
 325         }
 326         hammer_rel_mem_record(record);
 327 }
 328
 329 /*
 330  * Release a memory record.  Records marked for deletion are immediately
 331  * removed from the RB-Tree but otherwise left intact until the last ref
 332  * goes away.
 333  */
 334 void
 335 hammer_rel_mem_record(struct hammer_record *record)
 336 {
 337         hammer_inode_t ip, target_ip;
 338
 339         hammer_unref(&record->lock);
 340
 341         if (record->lock.refs == 0) {
 342                 /*
 343                  * Upon release of the last reference wakeup any waiters.
 344                  * The record structure may get destroyed so callers will
 345                  * loop up and do a relookup.
 346                  *
 347                  * WARNING!  Record must be removed from RB-TREE before we
 348                  * might possibly block.  hammer_test_inode() can block!
 349                  */
 350                 ip = record->ip;
 351
 352                 /*
 353                  * Upon release of the last reference a record marked deleted
 354                  * is destroyed.
 355                  */
 356                 if (record->flags & HAMMER_RECF_DELETED_FE) {
 357                         KKASSERT(ip->lock.refs > 0);
 358                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 359
 360                         /*
 361                          * target_ip may have zero refs, we have to ref it
 362                          * to prevent it from being ripped out from under
 363                          * us.
 364                          */
 365                         if ((target_ip = record->target_ip) != NULL) {
 366                                 TAILQ_REMOVE(&target_ip->target_list,
 367                                              record, target_entry);
 368                                 record->target_ip = NULL;
 369                                 hammer_ref(&target_ip->lock);
 370                         }
 371
 372                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 373                                 RB_REMOVE(hammer_rec_rb_tree,
 374                                           &record->ip->rec_tree,
 375                                           record);
 376                                 KKASSERT(ip->rsv_recs > 0);
 377                                 --ip->hmp->rsv_recs;
 378                                 --ip->rsv_recs;
 379                                 ip->hmp->rsv_databytes -= record->leaf.data_len;
 380                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 381
 382                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 383                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 384                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 385                                         hammer_test_inode(record->ip);
 386                                 }
 387                         }
 388
 389                         /*
 390                          * Do this test after removing record from the B-Tree.
 391                          */
 392                         if (target_ip) {
 393                                 hammer_test_inode(target_ip);
 394                                 hammer_rel_inode(target_ip, 0);
 395                         }
 396
 397                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 398                                 --hammer_count_record_datas;
 399                                 kfree(record->data, M_HAMMER);
 400                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 401                         }
 402                         if (record->resv) {
 403                                 hammer_blockmap_reserve_complete(ip->hmp,
 404                                                                  record->resv);
 405                                 record->resv = NULL;
 406                         }
 407                         record->data = NULL;
 408                         --hammer_count_records;
 409                         kfree(record, M_HAMMER);
 410                 }
 411         }
 412 }
 413
 414 /*
 415  * Record visibility depends on whether the record is being accessed by
 416  * the backend or the frontend.
 417  *
 418  * Return non-zero if the record is visible, zero if it isn't or if it is
 419  * deleted.
 420  */
 421 static __inline
 422 int
 423 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 424 {
 425         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 426                 if (record->flags & HAMMER_RECF_DELETED_BE)
 427                         return(0);
 428         } else {
 429                 if (record->flags & HAMMER_RECF_DELETED_FE)
 430                         return(0);
 431         }
 432         return(1);
 433 }
 434
 435 /*
 436  * This callback is used as part of the RB_SCAN function for in-memory
 437  * records.  We terminate it (return -1) as soon as we get a match.
 438  *
 439  * This routine is used by frontend code.
 440  *
 441  * The primary compare code does not account for ASOF lookups.  This
 442  * code handles that case as well as a few others.
 443  */
 444 static
 445 int
 446 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 447 {
 448         hammer_cursor_t cursor = data;
 449
 450         /*
 451          * We terminate on success, so this should be NULL on entry.
 452          */
 453         KKASSERT(cursor->iprec == NULL);
 454
 455         /*
 456          * Skip if the record was marked deleted.
 457          */
 458         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 459                 return(0);
 460
 461         /*
 462          * Skip if not visible due to our as-of TID
 463          */
 464         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 465                 if (cursor->asof < rec->leaf.base.create_tid)
 466                         return(0);
 467                 if (rec->leaf.base.delete_tid &&
 468                     cursor->asof >= rec->leaf.base.delete_tid) {
 469                         return(0);
 470                 }
 471         }
 472
 473         /*
 474          * If the record is queued to the flusher we have to block until
 475          * it isn't.  Otherwise we may see duplication between our memory
 476          * cache and the media.
 477          */
 478         hammer_ref(&rec->lock);
 479
 480 #warning "This deadlocks"
 481 #if 0
 482         if (rec->flush_state == HAMMER_FST_FLUSH)
 483                 hammer_wait_mem_record(rec);
 484 #endif
 485
 486         /*
 487          * The record may have been deleted while we were blocked.
 488          */
 489         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 490                 hammer_rel_mem_record(rec);
 491                 return(0);
 492         }
 493
 494         /*
 495          * Set the matching record and stop the scan.
 496          */
 497         cursor->iprec = rec;
 498         return(-1);
 499 }
 500
 501
 502 /*
 503  * Lookup an in-memory record given the key specified in the cursor.  Works
 504  * just like hammer_btree_lookup() but operates on an inode's in-memory
 505  * record list.
 506  *
 507  * The lookup must fail if the record is marked for deferred deletion.
 508  */
 509 static
 510 int
 511 hammer_mem_lookup(hammer_cursor_t cursor)
 512 {
 513         int error;
 514
 515         KKASSERT(cursor->ip);
 516         if (cursor->iprec) {
 517                 hammer_rel_mem_record(cursor->iprec);
 518                 cursor->iprec = NULL;
 519         }
 520         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 521                                    hammer_rec_scan_callback, cursor);
 522
 523         if (cursor->iprec == NULL)
 524                 error = ENOENT;
 525         else
 526                 error = 0;
 527         return(error);
 528 }
 529
 530 /*
 531  * hammer_mem_first() - locate the first in-memory record matching the
 532  * cursor within the bounds of the key range.
 533  */
 534 static
 535 int
 536 hammer_mem_first(hammer_cursor_t cursor)
 537 {
 538         hammer_inode_t ip;
 539
 540         ip = cursor->ip;
 541         KKASSERT(ip != NULL);
 542
 543         if (cursor->iprec) {
 544                 hammer_rel_mem_record(cursor->iprec);
 545                 cursor->iprec = NULL;
 546         }
 547
 548         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 549                                    hammer_rec_scan_callback, cursor);
 550
 551         /*
 552          * Adjust scan.node and keep it linked into the RB-tree so we can
 553          * hold the cursor through third party modifications of the RB-tree.
 554          */
 555         if (cursor->iprec)
 556                 return(0);
 557         return(ENOENT);
 558 }
 559
 560 void
 561 hammer_mem_done(hammer_cursor_t cursor)
 562 {
 563         if (cursor->iprec) {
 564                 hammer_rel_mem_record(cursor->iprec);
 565                 cursor->iprec = NULL;
 566         }
 567 }
 568
 569 /************************************************************************
 570  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 571  ************************************************************************
 572  *
 573  * These functions manipulate in-memory records.  Such records typically
 574  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 575  */
 576
 577 /*
 578  * Add a directory entry (dip,ncp) which references inode (ip).
 579  *
 580  * Note that the low 32 bits of the namekey are set temporarily to create
 581  * a unique in-memory record, and may be modified a second time when the
 582  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 583  * all 0's when synching to disk, which is not handled here.
 584  *
 585  * NOTE: bytes does not include any terminating \0 on name, and name might
 586  * not be terminated.
 587  */
 588 int
 589 hammer_ip_add_directory(struct hammer_transaction *trans,
 590                      struct hammer_inode *dip, const char *name, int bytes,
 591                      struct hammer_inode *ip)
 592 {
 593         hammer_record_t record;
 594         int error;
 595
 596         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 597         if (++trans->hmp->namekey_iterator == 0)
 598                 ++trans->hmp->namekey_iterator;
 599
 600         record->type = HAMMER_MEM_RECORD_ADD;
 601         record->leaf.base.localization = dip->obj_localization +
 602                                          HAMMER_LOCALIZE_MISC;
 603         record->leaf.base.obj_id = dip->obj_id;
 604         record->leaf.base.key = hammer_directory_namekey(name, bytes);
 605         record->leaf.base.key += trans->hmp->namekey_iterator;
 606         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 607         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 608         record->data->entry.obj_id = ip->obj_id;
 609         record->data->entry.localization = ip->obj_localization;
 610         bcopy(name, record->data->entry.name, bytes);
 611
 612         ++ip->ino_data.nlinks;
 613         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 614
 615         /*
 616          * The target inode and the directory entry are bound together.
 617          */
 618         record->target_ip = ip;
 619         record->flush_state = HAMMER_FST_SETUP;
 620         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 621
 622         /*
 623          * The inode now has a dependancy and must be taken out of the idle
 624          * state.  An inode not in an idle state is given an extra reference.
 625          */
 626         if (ip->flush_state == HAMMER_FST_IDLE) {
 627                 hammer_ref(&ip->lock);
 628                 ip->flush_state = HAMMER_FST_SETUP;
 629         }
 630         error = hammer_mem_add(record);
 631         return(error);
 632 }
 633
 634 /*
 635  * Delete the directory entry and update the inode link count.  The
 636  * cursor must be seeked to the directory entry record being deleted.
 637  *
 638  * The related inode should be share-locked by the caller.  The caller is
 639  * on the frontend.
 640  *
 641  * This function can return EDEADLK requiring the caller to terminate
 642  * the cursor, any locks, wait on the returned record, and retry.
 643  */
 644 int
 645 hammer_ip_del_directory(struct hammer_transaction *trans,
 646                      hammer_cursor_t cursor, struct hammer_inode *dip,
 647                      struct hammer_inode *ip)
 648 {
 649         hammer_record_t record;
 650         int error;
 651
 652         if (hammer_cursor_inmem(cursor)) {
 653                 /*
 654                  * In-memory (unsynchronized) records can simply be freed.
 655                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 656                  * by the backend, we must still avoid races against the
 657                  * backend potentially syncing the record to the media.
 658                  *
 659                  * We cannot call hammer_ip_delete_record(), that routine may
 660                  * only be called from the backend.
 661                  */
 662                 record = cursor->iprec;
 663                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 664                         KKASSERT(cursor->deadlk_rec == NULL);
 665                         hammer_ref(&record->lock);
 666                         cursor->deadlk_rec = record;
 667                         error = EDEADLK;
 668                 } else {
 669                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 670                         record->flags |= HAMMER_RECF_DELETED_FE;
 671                         error = 0;
 672                 }
 673         } else {
 674                 /*
 675                  * If the record is on-disk we have to queue the deletion by
 676                  * the record's key.  This also causes lookups to skip the
 677                  * record.
 678                  */
 679                 KKASSERT(dip->flags &
 680                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 681                 record = hammer_alloc_mem_record(dip, 0);
 682                 record->type = HAMMER_MEM_RECORD_DEL;
 683                 record->leaf.base = cursor->leaf->base;
 684
 685                 record->target_ip = ip;
 686                 record->flush_state = HAMMER_FST_SETUP;
 687                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 688
 689                 /*
 690                  * The inode now has a dependancy and must be taken out of
 691                  * the idle state.  An inode not in an idle state is given
 692                  * an extra reference.
 693                  */
 694                 if (ip->flush_state == HAMMER_FST_IDLE) {
 695                         hammer_ref(&ip->lock);
 696                         ip->flush_state = HAMMER_FST_SETUP;
 697                 }
 698
 699                 error = hammer_mem_add(record);
 700         }
 701
 702         /*
 703          * One less link.  The file may still be open in the OS even after
 704          * all links have gone away.
 705          *
 706          * We have to terminate the cursor before syncing the inode to
 707          * avoid deadlocking against ourselves.  XXX this may no longer
 708          * be true.
 709          *
 710          * If nlinks drops to zero and the vnode is inactive (or there is
 711          * no vnode), call hammer_inode_unloadable_check() to zonk the
 712          * inode.  If we don't do this here the inode will not be destroyed
 713          * on-media until we unmount.
 714          */
 715         if (error == 0) {
 716                 --ip->ino_data.nlinks;
 717                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 718                 if (ip->ino_data.nlinks == 0 &&
 719                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 720                         hammer_done_cursor(cursor);
 721                         hammer_inode_unloadable_check(ip, 1);
 722                         hammer_flush_inode(ip, 0);
 723                 }
 724
 725         }
 726         return(error);
 727 }
 728
 729 /*
 730  * Add a record to an inode.
 731  *
 732  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 733  * initialize the following additional fields:
 734  *
 735  * The related inode should be share-locked by the caller.  The caller is
 736  * on the frontend.
 737  *
 738  * record->rec.entry.base.base.key
 739  * record->rec.entry.base.base.rec_type
 740  * record->rec.entry.base.base.data_len
 741  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 742  */
 743 int
 744 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 745 {
 746         hammer_inode_t ip = record->ip;
 747         int error;
 748
 749         KKASSERT(record->leaf.base.localization != 0);
 750         record->leaf.base.obj_id = ip->obj_id;
 751         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 752         error = hammer_mem_add(record);
 753         return(error);
 754 }
 755
 756 /*
 757  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 758  * reserved so the front-end can flush large data writes without having
 759  * to queue the BIO to the flusher.  Only the related record gets queued
 760  * to the flusher.
 761  */
 762 static hammer_record_t
 763 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 764 {
 765         hammer_record_t record;
 766         struct hammer_btree_leaf_elm leaf;
 767
 768         bzero(&leaf, sizeof(leaf));
 769         leaf.base.obj_id = ip->obj_id;
 770         leaf.base.key = file_offset + bytes;
 771         leaf.base.create_tid = 0;
 772         leaf.base.delete_tid = 0;
 773         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 774         leaf.base.obj_type = 0;                 /* unused */
 775         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
 776         leaf.base.localization = ip->obj_localization + HAMMER_LOCALIZE_MISC;
 777         leaf.data_len = bytes;
 778
 779         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
 780         if (record)
 781                 hammer_ref(&record->lock);
 782         return(record);
 783 }
 784
 785 /*
 786  * Reserve blockmap space placemarked with an in-memory record.
 787  *
 788  * This routine is called by the frontend in order to be able to directly
 789  * flush a buffer cache buffer.  The frontend has locked the related buffer
 790  * cache buffers and we should be able to manipulate any overlapping
 791  * in-memory records.
 792  */
 793 hammer_record_t
 794 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 795                    int *errorp)
 796 {
 797         hammer_record_t record;
 798         hammer_record_t conflict;
 799         int zone;
 800         int flags;
 801
 802         /*
 803          * Deal with conflicting in-memory records.  We cannot have multiple
 804          * in-memory records for the same offset without seriously confusing
 805          * the backend, including but not limited to the backend issuing
 806          * delete-create-delete sequences and asserting on the delete_tid
 807          * being the same as the create_tid.
 808          *
 809          * If we encounter a record with the backend interlock set we cannot
 810          * immediately delete it without confusing the backend.
 811          */
 812         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 813                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 814                         conflict->flags |= HAMMER_RECF_WANTED;
 815                         tsleep(conflict, 0, "hmrrc3", 0);
 816                 } else {
 817                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 818                 }
 819                 hammer_rel_mem_record(conflict);
 820         }
 821
 822         /*
 823          * Create a record to cover the direct write.  This is called with
 824          * the related BIO locked so there should be no possible conflict.
 825          *
 826          * The backend is responsible for finalizing the space reserved in
 827          * this record.
 828          *
 829          * XXX bytes not aligned, depend on the reservation code to
 830          * align the reservation.
 831          */
 832         record = hammer_alloc_mem_record(ip, 0);
 833         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 834                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 835         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 836                                                &record->leaf.data_offset,
 837                                                errorp);
 838         if (record->resv == NULL) {
 839                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 840                 hammer_rel_mem_record(record);
 841                 return(NULL);
 842         }
 843         record->type = HAMMER_MEM_RECORD_DATA;
 844         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 845         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 846         record->leaf.base.obj_id = ip->obj_id;
 847         record->leaf.base.key = file_offset + bytes;
 848         record->leaf.base.localization = ip->obj_localization +
 849                                          HAMMER_LOCALIZE_MISC;
 850         record->leaf.data_len = bytes;
 851         hammer_crc_set_leaf(data, &record->leaf);
 852         flags = record->flags;
 853
 854         hammer_ref(&record->lock);      /* mem_add eats a reference */
 855         *errorp = hammer_mem_add(record);
 856         if (*errorp) {
 857                 conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 858                 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
 859                         *errorp, conflict, file_offset, bytes);
 860                 if (conflict)
 861                         kprintf("conflict %lld %d\n", conflict->leaf.base.key, conflict->leaf.data_len);
 862                 if (conflict)
 863                         hammer_rel_mem_record(conflict);
 864         }
 865         KKASSERT(*errorp == 0);
 866         conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 867         if (conflict != record) {
 868                 kprintf("conflict mismatch %p %p %08x\n", conflict, record, record->flags);
 869                 if (conflict)
 870                     kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict->leaf.base.key, conflict->leaf.data_len, record->leaf.base.key, record->leaf.data_len);
 871         }
 872         KKASSERT(conflict == record);
 873         hammer_rel_mem_record(conflict);
 874
 875         return (record);
 876 }
 877
 878 /*
 879  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 880  * and records in a flushing state are handled by the backend.  The vnops
 881  * setattr code will handle the block containing the truncation point.
 882  *
 883  * Partial blocks are not deleted.
 884  */
 885 int
 886 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 887 {
 888         struct rec_trunc_info info;
 889
 890         switch(ip->ino_data.obj_type) {
 891         case HAMMER_OBJTYPE_REGFILE:
 892                 info.rec_type = HAMMER_RECTYPE_DATA;
 893                 break;
 894         case HAMMER_OBJTYPE_DBFILE:
 895                 info.rec_type = HAMMER_RECTYPE_DB;
 896                 break;
 897         default:
 898                 return(EINVAL);
 899         }
 900         info.trunc_off = file_size;
 901         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 902                                    hammer_rec_trunc_callback, &info);
 903         return(0);
 904 }
 905
 906 static int
 907 hammer_rec_trunc_callback(hammer_record_t record, void *data __unused)
 908 {
 909         if (record->flags & HAMMER_RECF_DELETED_FE)
 910                 return(0);
 911         if (record->flush_state == HAMMER_FST_FLUSH)
 912                 return(0);
 913         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 914         hammer_ref(&record->lock);
 915         record->flags |= HAMMER_RECF_DELETED_FE;
 916         hammer_rel_mem_record(record);
 917         return(0);
 918 }
 919
 920 /*
 921  * Return 1 if the caller must check for and delete existing records
 922  * before writing out a new data record.
 923  *
 924  * Return 0 if the caller can just insert the record into the B-Tree without
 925  * checking.
 926  */
 927 static int
 928 hammer_record_needs_overwrite_delete(hammer_record_t record)
 929 {
 930         hammer_inode_t ip = record->ip;
 931         int64_t file_offset;
 932         int r;
 933
 934         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
 935                 file_offset = record->leaf.base.key;
 936         else
 937                 file_offset = record->leaf.base.key - record->leaf.data_len;
 938         r = (file_offset < ip->sync_trunc_off);
 939         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
 940                 if (ip->sync_trunc_off <= record->leaf.base.key)
 941                         ip->sync_trunc_off = record->leaf.base.key + 1;
 942         } else {
 943                 if (ip->sync_trunc_off < record->leaf.base.key)
 944                         ip->sync_trunc_off = record->leaf.base.key;
 945         }
 946         return(r);
 947 }
 948
 949 /*
 950  * Backend code.  Sync a record to the media.
 951  */
 952 int
 953 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
 954 {
 955         hammer_transaction_t trans = cursor->trans;
 956         int64_t file_offset;
 957         int bytes;
 958         void *bdata;
 959         int error;
 960
 961         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 962         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 963         KKASSERT(record->leaf.base.localization != 0);
 964
 965         /*
 966          * If this is a bulk-data record placemarker there may be an existing
 967          * record on-disk, indicating a data overwrite.  If there is the
 968          * on-disk record must be deleted before we can insert our new record.
 969          *
 970          * We've synthesized this record and do not know what the create_tid
 971          * on-disk is, nor how much data it represents.
 972          *
 973          * Keep in mind that (key) for data records is (base_offset + len),
 974          * not (base_offset).  Also, we only want to get rid of on-disk
 975          * records since we are trying to sync our in-memory record, call
 976          * hammer_ip_delete_range() with truncating set to 1 to make sure
 977          * it skips in-memory records.
 978          *
 979          * It is ok for the lookup to return ENOENT.
 980          *
 981          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
 982          * to call hammer_ip_delete_range() or not.  This also means we must
 983          * update sync_trunc_off() as we write.
 984          */
 985         if (record->type == HAMMER_MEM_RECORD_DATA &&
 986             hammer_record_needs_overwrite_delete(record)) {
 987                 file_offset = record->leaf.base.key - record->leaf.data_len;
 988                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
 989                         ~HAMMER_BUFMASK;
 990                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
 991                 error = hammer_ip_delete_range(
 992                                 cursor, record->ip,
 993                                 file_offset, file_offset + bytes - 1,
 994                                 1);
 995                 if (error && error != ENOENT)
 996                         goto done;
 997         }
 998
 999         /*
1000          * Setup the cursor.
1001          */
1002         hammer_normalize_cursor(cursor);
1003         cursor->key_beg = record->leaf.base;
1004         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1005         cursor->flags |= HAMMER_CURSOR_BACKEND;
1006         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1007
1008         /*
1009          * Records can wind up on-media before the inode itself is on-media.
1010          * Flag the case.
1011          */
1012         record->ip->flags |= HAMMER_INODE_DONDISK;
1013
1014         /*
1015          * If we are deleting a directory entry an exact match must be
1016          * found on-disk.
1017          */
1018         if (record->type == HAMMER_MEM_RECORD_DEL) {
1019                 error = hammer_btree_lookup(cursor);
1020                 if (error == 0) {
1021                         error = hammer_ip_delete_record(cursor, record->ip,
1022                                                         trans->tid);
1023                         if (error == 0) {
1024                                 record->flags |= HAMMER_RECF_DELETED_FE;
1025                                 record->flags |= HAMMER_RECF_DELETED_BE;
1026                         }
1027                 }
1028                 goto done;
1029         }
1030
1031         /*
1032          * We are inserting.
1033          *
1034          * Issue a lookup to position the cursor and locate the cluster.  The
1035          * target key should not exist.  If we are creating a directory entry
1036          * we may have to iterate the low 32 bits of the key to find an unused
1037          * key.
1038          */
1039         cursor->flags |= HAMMER_CURSOR_INSERT;
1040
1041         for (;;) {
1042                 error = hammer_btree_lookup(cursor);
1043                 if (hammer_debug_inode)
1044                         kprintf("DOINSERT LOOKUP %d\n", error);
1045                 if (error)
1046                         break;
1047                 if (record->leaf.base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
1048                         kprintf("hammer_ip_sync_record: duplicate rec "
1049                                 "at (%016llx)\n", record->leaf.base.key);
1050                         Debugger("duplicate record1");
1051                         error = EIO;
1052                         break;
1053                 }
1054                 if (++trans->hmp->namekey_iterator == 0)
1055                         ++trans->hmp->namekey_iterator;
1056                 record->leaf.base.key &= ~(0xFFFFFFFFLL);
1057                 record->leaf.base.key |= trans->hmp->namekey_iterator;
1058                 cursor->key_beg.key = record->leaf.base.key;
1059         }
1060 #if 0
1061         if (record->type == HAMMER_MEM_RECORD_DATA)
1062                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1063                         record->leaf.base.key - record->leaf.data_len,
1064                         record->leaf.data_offset, error);
1065 #endif
1066
1067
1068         if (error != ENOENT)
1069                 goto done;
1070
1071         /*
1072          * Allocate the record and data.  The result buffers will be
1073          * marked as being modified and further calls to
1074          * hammer_modify_buffer() will result in unneeded UNDO records.
1075          *
1076          * Support zero-fill records (data == NULL and data_len != 0)
1077          */
1078         if (record->type == HAMMER_MEM_RECORD_DATA) {
1079                 /*
1080                  * The data portion of a bulk-data record has already been
1081                  * committed to disk, we need only adjust the layer2
1082                  * statistics in the same transaction as our B-Tree insert.
1083                  */
1084                 KKASSERT(record->leaf.data_offset != 0);
1085                 hammer_blockmap_finalize(trans, record->leaf.data_offset,
1086                                          record->leaf.data_len);
1087                 error = 0;
1088         } else if (record->data && record->leaf.data_len) {
1089                 /*
1090                  * Wholely cached record, with data.  Allocate the data.
1091                  */
1092                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1093                                           record->leaf.base.rec_type,
1094                                           &record->leaf.data_offset,
1095                                           &cursor->data_buffer, &error);
1096                 if (bdata == NULL)
1097                         goto done;
1098                 hammer_crc_set_leaf(record->data, &record->leaf);
1099                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1100                 bcopy(record->data, bdata, record->leaf.data_len);
1101                 hammer_modify_buffer_done(cursor->data_buffer);
1102         } else {
1103                 /*
1104                  * Wholely cached record, without data.
1105                  */
1106                 record->leaf.data_offset = 0;
1107                 record->leaf.data_crc = 0;
1108         }
1109
1110         error = hammer_btree_insert(cursor, &record->leaf);
1111         if (hammer_debug_inode && error)
1112                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1113
1114         /*
1115          * Our record is on-disk, normally mark the in-memory version as
1116          * deleted.  If the record represented a directory deletion but
1117          * we had to sync a valid directory entry to disk we must convert
1118          * the record to a covering delete so the frontend does not have
1119          * visibility on the synced entry.
1120          */
1121         if (error == 0) {
1122                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1123                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1124                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1125                         record->type = HAMMER_MEM_RECORD_DEL;
1126                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1127                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1128                         /* hammer_flush_record_done takes care of the rest */
1129                 } else {
1130                         record->flags |= HAMMER_RECF_DELETED_FE;
1131                         record->flags |= HAMMER_RECF_DELETED_BE;
1132                 }
1133         } else {
1134                 if (record->leaf.data_offset) {
1135                         hammer_blockmap_free(trans, record->leaf.data_offset,
1136                                              record->leaf.data_len);
1137                 }
1138         }
1139
1140 done:
1141         return(error);
1142 }
1143
1144 /*
1145  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1146  * entry's key is used to deal with hash collisions in the upper 32 bits.
1147  * A unique 64 bit key is generated in-memory and may be regenerated a
1148  * second time when the directory record is flushed to the on-disk B-Tree.
1149  *
1150  * A referenced record is passed to this function.  This function
1151  * eats the reference.  If an error occurs the record will be deleted.
1152  *
1153  * A copy of the temporary record->data pointer provided by the caller
1154  * will be made.
1155  */
1156 static
1157 int
1158 hammer_mem_add(hammer_record_t record)
1159 {
1160         hammer_mount_t hmp = record->ip->hmp;
1161
1162         /*
1163          * Make a private copy of record->data
1164          */
1165         if (record->data)
1166                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1167
1168         /*
1169          * Insert into the RB tree, find an unused iterator if this is
1170          * a directory entry.
1171          */
1172         while (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1173                 if (record->leaf.base.rec_type != HAMMER_RECTYPE_DIRENTRY){
1174                         record->flags |= HAMMER_RECF_DELETED_FE;
1175                         hammer_rel_mem_record(record);
1176                         return (EEXIST);
1177                 }
1178                 if (++hmp->namekey_iterator == 0)
1179                         ++hmp->namekey_iterator;
1180                 record->leaf.base.key &= ~(0xFFFFFFFFLL);
1181                 record->leaf.base.key |= hmp->namekey_iterator;
1182         }
1183         ++hmp->count_newrecords;
1184         ++hmp->rsv_recs;
1185         ++record->ip->rsv_recs;
1186         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1187         record->flags |= HAMMER_RECF_ONRBTREE;
1188         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1189         hammer_rel_mem_record(record);
1190         return(0);
1191 }
1192
1193 /************************************************************************
1194  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1195  ************************************************************************
1196  *
1197  * These functions augment the B-Tree scanning functions in hammer_btree.c
1198  * by merging in-memory records with on-disk records.
1199  */
1200
1201 /*
1202  * Locate a particular record either in-memory or on-disk.
1203  *
1204  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1205  * NOT be called to iterate results.
1206  */
1207 int
1208 hammer_ip_lookup(hammer_cursor_t cursor)
1209 {
1210         int error;
1211
1212         /*
1213          * If the element is in-memory return it without searching the
1214          * on-disk B-Tree
1215          */
1216         KKASSERT(cursor->ip);
1217         error = hammer_mem_lookup(cursor);
1218         if (error == 0) {
1219                 cursor->leaf = &cursor->iprec->leaf;
1220                 return(error);
1221         }
1222         if (error != ENOENT)
1223                 return(error);
1224
1225         /*
1226          * If the inode has on-disk components search the on-disk B-Tree.
1227          */
1228         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1229                 return(error);
1230         error = hammer_btree_lookup(cursor);
1231         if (error == 0)
1232                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1233         return(error);
1234 }
1235
1236 /*
1237  * Locate the first record within the cursor's key_beg/key_end range,
1238  * restricted to a particular inode.  0 is returned on success, ENOENT
1239  * if no records matched the requested range, or some other error.
1240  *
1241  * When 0 is returned hammer_ip_next() may be used to iterate additional
1242  * records within the requested range.
1243  *
1244  * This function can return EDEADLK, requiring the caller to terminate
1245  * the cursor and try again.
1246  */
1247 int
1248 hammer_ip_first(hammer_cursor_t cursor)
1249 {
1250         hammer_inode_t ip = cursor->ip;
1251         int error;
1252
1253         KKASSERT(ip != NULL);
1254
1255         /*
1256          * Clean up fields and setup for merged scan
1257          */
1258         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1259         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1260         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1261         if (cursor->iprec) {
1262                 hammer_rel_mem_record(cursor->iprec);
1263                 cursor->iprec = NULL;
1264         }
1265
1266         /*
1267          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1268          * exact lookup so if we get ENOENT we have to call the iterate
1269          * function to validate the first record after the begin key.
1270          *
1271          * The ATEDISK flag is used by hammer_btree_iterate to determine
1272          * whether it must index forwards or not.  It is also used here
1273          * to select the next record from in-memory or on-disk.
1274          *
1275          * EDEADLK can only occur if the lookup hit an empty internal
1276          * element and couldn't delete it.  Since this could only occur
1277          * in-range, we can just iterate from the failure point.
1278          */
1279         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1280                 error = hammer_btree_lookup(cursor);
1281                 if (error == ENOENT || error == EDEADLK) {
1282                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1283                         if (hammer_debug_general & 0x2000)
1284                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1285                         error = hammer_btree_iterate(cursor);
1286                 }
1287                 if (error && error != ENOENT)
1288                         return(error);
1289                 if (error == 0) {
1290                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1291                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1292                 } else {
1293                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1294                 }
1295         }
1296
1297         /*
1298          * Search the in-memory record list (Red-Black tree).  Unlike the
1299          * B-Tree search, mem_first checks for records in the range.
1300          */
1301         error = hammer_mem_first(cursor);
1302         if (error && error != ENOENT)
1303                 return(error);
1304         if (error == 0) {
1305                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1306                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1307                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1308                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1309         }
1310
1311         /*
1312          * This will return the first matching record.
1313          */
1314         return(hammer_ip_next(cursor));
1315 }
1316
1317 /*
1318  * Retrieve the next record in a merged iteration within the bounds of the
1319  * cursor.  This call may be made multiple times after the cursor has been
1320  * initially searched with hammer_ip_first().
1321  *
1322  * 0 is returned on success, ENOENT if no further records match the
1323  * requested range, or some other error code is returned.
1324  */
1325 int
1326 hammer_ip_next(hammer_cursor_t cursor)
1327 {
1328         hammer_btree_elm_t elm;
1329         hammer_record_t rec, save;
1330         int error;
1331         int r;
1332
1333 next_btree:
1334         /*
1335          * Load the current on-disk and in-memory record.  If we ate any
1336          * records we have to get the next one.
1337          *
1338          * If we deleted the last on-disk record we had scanned ATEDISK will
1339          * be clear and DELBTREE will be set, forcing a call to iterate. The
1340          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1341          * element.  If ATEDISK is set, iterate will skip the 'current'
1342          * element.
1343          *
1344          * Get the next on-disk record
1345          */
1346         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1347                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1348                         error = hammer_btree_iterate(cursor);
1349                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1350                         if (error == 0) {
1351                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1352                                 hammer_cache_node(&cursor->ip->cache[1],
1353                                                   cursor->node);
1354                         } else {
1355                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1356                                                  HAMMER_CURSOR_ATEDISK;
1357                         }
1358                 }
1359         }
1360
1361 next_memory:
1362         /*
1363          * Get the next in-memory record.  The record can be ripped out
1364          * of the RB tree so we maintain a scan_info structure to track
1365          * the next node.
1366          *
1367          * hammer_rec_scan_cmp:  Is the record still in our general range,
1368          *                       (non-inclusive of snapshot exclusions)?
1369          * hammer_rec_scan_callback: Is the record in our snapshot?
1370          */
1371         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1372                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1373                         save = cursor->iprec;
1374                         cursor->iprec = NULL;
1375                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1376                         while (rec) {
1377                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1378                                         break;
1379                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1380                                         break;
1381                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1382                         }
1383                         if (save)
1384                                 hammer_rel_mem_record(save);
1385                         if (cursor->iprec) {
1386                                 KKASSERT(cursor->iprec == rec);
1387                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1388                         } else {
1389                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1390                         }
1391                 }
1392         }
1393
1394         /*
1395          * The memory record may have become stale while being held in
1396          * cursor->iprec.  We are interlocked against the backend on
1397          * with regards to B-Tree entries.
1398          */
1399         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1400                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1401                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1402                         goto next_memory;
1403                 }
1404         }
1405
1406         /*
1407          * Extract either the disk or memory record depending on their
1408          * relative position.
1409          */
1410         error = 0;
1411         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1412         case 0:
1413                 /*
1414                  * Both entries valid.   Compare the entries and nominally
1415                  * return the first one in the sort order.  Numerous cases
1416                  * require special attention, however.
1417                  */
1418                 elm = &cursor->node->ondisk->elms[cursor->index];
1419                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1420
1421                 /*
1422                  * If the two entries differ only by their key (-2/2) or
1423                  * create_tid (-1/1), and are DATA records, we may have a
1424                  * nominal match.  We have to calculate the base file
1425                  * offset of the data.
1426                  */
1427                 if (r <= 2 && r >= -2 && r != 0 &&
1428                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1429                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1430                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1431                         int64_t base2 = cursor->iprec->leaf.base.key -
1432                                         cursor->iprec->leaf.data_len;
1433                         if (base1 == base2)
1434                                 r = 0;
1435                 }
1436
1437                 if (r < 0) {
1438                         error = hammer_btree_extract(cursor,
1439                                                      HAMMER_CURSOR_GET_LEAF);
1440                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1441                         break;
1442                 }
1443
1444                 /*
1445                  * If the entries match exactly the memory entry is either
1446                  * an on-disk directory entry deletion or a bulk data
1447                  * overwrite.  If it is a directory entry deletion we eat
1448                  * both entries.
1449                  *
1450                  * For the bulk-data overwrite case it is possible to have
1451                  * visibility into both, which simply means the syncer
1452                  * hasn't gotten around to doing the delete+insert sequence
1453                  * on the B-Tree.  Use the memory entry and throw away the
1454                  * on-disk entry.
1455                  *
1456                  * If the in-memory record is not either of these we
1457                  * probably caught the syncer while it was syncing it to
1458                  * the media.  Since we hold a shared lock on the cursor,
1459                  * the in-memory record had better be marked deleted at
1460                  * this point.
1461                  */
1462                 if (r == 0) {
1463                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1464                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1465                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1466                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1467                                         goto next_btree;
1468                                 }
1469                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1470                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1471                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1472                                 }
1473                                 /* fall through to memory entry */
1474                         } else {
1475                                 panic("hammer_ip_next: duplicate mem/b-tree entry");
1476                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1477                                 goto next_memory;
1478                         }
1479                 }
1480                 /* fall through to the memory entry */
1481         case HAMMER_CURSOR_ATEDISK:
1482                 /*
1483                  * Only the memory entry is valid.
1484                  */
1485                 cursor->leaf = &cursor->iprec->leaf;
1486                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1487
1488                 /*
1489                  * If the memory entry is an on-disk deletion we should have
1490                  * also had found a B-Tree record.  If the backend beat us
1491                  * to it it would have interlocked the cursor and we should
1492                  * have seen the in-memory record marked DELETED_FE.
1493                  */
1494                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1495                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1496                         panic("hammer_ip_next: del-on-disk with no b-tree entry");
1497                 }
1498                 break;
1499         case HAMMER_CURSOR_ATEMEM:
1500                 /*
1501                  * Only the disk entry is valid
1502                  */
1503                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1504                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1505                 break;
1506         default:
1507                 /*
1508                  * Neither entry is valid
1509                  *
1510                  * XXX error not set properly
1511                  */
1512                 cursor->leaf = NULL;
1513                 error = ENOENT;
1514                 break;
1515         }
1516         return(error);
1517 }
1518
1519 /*
1520  * Resolve the cursor->data pointer for the current cursor position in
1521  * a merged iteration.
1522  */
1523 int
1524 hammer_ip_resolve_data(hammer_cursor_t cursor)
1525 {
1526         hammer_record_t record;
1527         int error;
1528
1529         if (hammer_cursor_inmem(cursor)) {
1530                 /*
1531                  * The data associated with an in-memory record is usually
1532                  * kmalloced, but reserve-ahead data records will have an
1533                  * on-disk reference.
1534                  *
1535                  * NOTE: Reserve-ahead data records must be handled in the
1536                  * context of the related high level buffer cache buffer
1537                  * to interlock against async writes.
1538                  */
1539                 record = cursor->iprec;
1540                 cursor->data = record->data;
1541                 error = 0;
1542                 if (cursor->data == NULL) {
1543                         KKASSERT(record->leaf.base.rec_type ==
1544                                  HAMMER_RECTYPE_DATA);
1545                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1546                                                     record->leaf.data_offset,
1547                                                     record->leaf.data_len,
1548                                                     &error,
1549                                                     &cursor->data_buffer);
1550                 }
1551         } else {
1552                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1553                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1554         }
1555         return(error);
1556 }
1557
1558 /*
1559  * Backend truncation / record replacement - delete records in range.
1560  *
1561  * Delete all records within the specified range for inode ip.  In-memory
1562  * records still associated with the frontend are ignored.
1563  *
1564  * NOTE: An unaligned range will cause new records to be added to cover
1565  * the edge cases. (XXX not implemented yet).
1566  *
1567  * NOTE: Replacement via reservations (see hammer_ip_sync_record_cursor())
1568  * also do not deal with unaligned ranges.
1569  *
1570  * NOTE: ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1571  *
1572  * NOTE: Record keys for regular file data have to be special-cased since
1573  * they indicate the end of the range (key = base + bytes).
1574  */
1575 int
1576 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1577                        int64_t ran_beg, int64_t ran_end, int truncating)
1578 {
1579         hammer_transaction_t trans = cursor->trans;
1580         hammer_btree_leaf_elm_t leaf;
1581         int error;
1582         int64_t off;
1583
1584 #if 0
1585         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1586 #endif
1587
1588         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1589 retry:
1590         hammer_normalize_cursor(cursor);
1591         cursor->key_beg.localization = ip->obj_localization +
1592                                        HAMMER_LOCALIZE_MISC;
1593         cursor->key_beg.obj_id = ip->obj_id;
1594         cursor->key_beg.create_tid = 0;
1595         cursor->key_beg.delete_tid = 0;
1596         cursor->key_beg.obj_type = 0;
1597         cursor->asof = ip->obj_asof;
1598         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1599         cursor->flags |= HAMMER_CURSOR_ASOF;
1600         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1601         cursor->flags |= HAMMER_CURSOR_BACKEND;
1602
1603         cursor->key_end = cursor->key_beg;
1604         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1605                 cursor->key_beg.key = ran_beg;
1606                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1607                 cursor->key_end.rec_type = HAMMER_RECTYPE_DB;
1608                 cursor->key_end.key = ran_end;
1609         } else {
1610                 /*
1611                  * The key in the B-Tree is (base+bytes), so the first possible
1612                  * matching key is ran_beg + 1.
1613                  */
1614                 int64_t tmp64;
1615
1616                 cursor->key_beg.key = ran_beg + 1;
1617                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1618                 cursor->key_end.rec_type = HAMMER_RECTYPE_DATA;
1619
1620                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1621                 if (tmp64 < ran_end)
1622                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1623                 else
1624                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1625         }
1626         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1627
1628         error = hammer_ip_first(cursor);
1629
1630         /*
1631          * Iterate through matching records and mark them as deleted.
1632          */
1633         while (error == 0) {
1634                 leaf = cursor->leaf;
1635
1636                 KKASSERT(leaf->base.delete_tid == 0);
1637
1638                 /*
1639                  * There may be overlap cases for regular file data.  Also
1640                  * remember the key for a regular file record is (base + len),
1641                  * NOT (base).
1642                  */
1643                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1644                         off = leaf->base.key - leaf->data_len;
1645                         /*
1646                          * Check the left edge case.  We currently do not
1647                          * split existing records.
1648                          */
1649                         if (off < ran_beg) {
1650                                 panic("hammer left edge case %016llx %d\n",
1651                                         leaf->base.key, leaf->data_len);
1652                         }
1653
1654                         /*
1655                          * Check the right edge case.  Note that the
1656                          * record can be completely out of bounds, which
1657                          * terminates the search.
1658                          *
1659                          * base->key is exclusive of the right edge while
1660                          * ran_end is inclusive of the right edge.  The
1661                          * (key - data_len) left boundary is inclusive.
1662                          *
1663                          * XXX theory-check this test at some point, are
1664                          * we missing a + 1 somewhere?  Note that ran_end
1665                          * could overflow.
1666                          */
1667                         if (leaf->base.key - 1 > ran_end) {
1668                                 if (leaf->base.key - leaf->data_len > ran_end)
1669                                         break;
1670                                 panic("hammer right edge case\n");
1671                         }
1672                 }
1673
1674                 /*
1675                  * Delete the record.  When truncating we do not delete
1676                  * in-memory (data) records because they represent data
1677                  * written after the truncation.
1678                  *
1679                  * This will also physically destroy the B-Tree entry and
1680                  * data if the retention policy dictates.  The function
1681                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1682                  * uses to perform a fixup.
1683                  */
1684                 if (truncating == 0 || hammer_cursor_ondisk(cursor))
1685                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1686                 if (error)
1687                         break;
1688                 error = hammer_ip_next(cursor);
1689         }
1690         if (cursor->node)
1691                 hammer_cache_node(&ip->cache[1], cursor->node);
1692
1693         if (error == EDEADLK) {
1694                 hammer_done_cursor(cursor);
1695                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1696                 if (error == 0)
1697                         goto retry;
1698         }
1699         if (error == ENOENT)
1700                 error = 0;
1701         return(error);
1702 }
1703
1704 /*
1705  * Backend truncation - delete all records.
1706  *
1707  * Delete all user records associated with an inode except the inode record
1708  * itself.  Directory entries are not deleted (they must be properly disposed
1709  * of or nlinks would get upset).
1710  */
1711 int
1712 hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
1713                            int *countp)
1714 {
1715         hammer_transaction_t trans = cursor->trans;
1716         hammer_btree_leaf_elm_t leaf;
1717         int error;
1718
1719         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1720 retry:
1721         hammer_normalize_cursor(cursor);
1722         cursor->key_beg.localization = ip->obj_localization +
1723                                        HAMMER_LOCALIZE_MISC;
1724         cursor->key_beg.obj_id = ip->obj_id;
1725         cursor->key_beg.create_tid = 0;
1726         cursor->key_beg.delete_tid = 0;
1727         cursor->key_beg.obj_type = 0;
1728         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1729         cursor->key_beg.key = HAMMER_MIN_KEY;
1730
1731         cursor->key_end = cursor->key_beg;
1732         cursor->key_end.rec_type = 0xFFFF;
1733         cursor->key_end.key = HAMMER_MAX_KEY;
1734
1735         cursor->asof = ip->obj_asof;
1736         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1737         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1738         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1739         cursor->flags |= HAMMER_CURSOR_BACKEND;
1740
1741         error = hammer_ip_first(cursor);
1742
1743         /*
1744          * Iterate through matching records and mark them as deleted.
1745          */
1746         while (error == 0) {
1747                 leaf = cursor->leaf;
1748
1749                 KKASSERT(leaf->base.delete_tid == 0);
1750
1751                 /*
1752                  * Mark the record and B-Tree entry as deleted.  This will
1753                  * also physically delete the B-Tree entry, record, and
1754                  * data if the retention policy dictates.  The function
1755                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1756                  * uses to perform a fixup.
1757                  *
1758                  * Directory entries (and delete-on-disk directory entries)
1759                  * must be synced and cannot be deleted.
1760                  */
1761                 if (leaf->base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
1762                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1763                         ++*countp;
1764                 }
1765                 if (error)
1766                         break;
1767                 error = hammer_ip_next(cursor);
1768         }
1769         if (cursor->node)
1770                 hammer_cache_node(&ip->cache[1], cursor->node);
1771         if (error == EDEADLK) {
1772                 hammer_done_cursor(cursor);
1773                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1774                 if (error == 0)
1775                         goto retry;
1776         }
1777         if (error == ENOENT)
1778                 error = 0;
1779         return(error);
1780 }
1781
1782 /*
1783  * Delete the record at the current cursor.  On success the cursor will
1784  * be positioned appropriately for an iteration but may no longer be at
1785  * a leaf node.
1786  *
1787  * This routine is only called from the backend.
1788  *
1789  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1790  * cursor and retry.
1791  */
1792 int
1793 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1794                         hammer_tid_t tid)
1795 {
1796         hammer_off_t zone2_offset;
1797         hammer_record_t iprec;
1798         hammer_btree_elm_t elm;
1799         hammer_mount_t hmp;
1800         int error;
1801         int dodelete;
1802
1803         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1804         KKASSERT(tid != 0);
1805         hmp = cursor->node->hmp;
1806
1807         /*
1808          * In-memory (unsynchronized) records can simply be freed.  This
1809          * only occurs in range iterations since all other records are
1810          * individually synchronized.  Thus there should be no confusion with
1811          * the interlock.
1812          *
1813          * An in-memory record may be deleted before being committed to disk,
1814          * but could have been accessed in the mean time.  The backing store
1815          * may never been marked allocated and so hammer_blockmap_free() may
1816          * never get called on it.  Because of this we have to make sure that
1817          * we've gotten rid of any related hammer_buffer or buffer cache
1818          * buffer.
1819          */
1820         if (hammer_cursor_inmem(cursor)) {
1821                 iprec = cursor->iprec;
1822                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1823                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1824                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1825
1826                 if (iprec->leaf.data_offset && iprec->leaf.data_len) {
1827                         zone2_offset = hammer_blockmap_lookup(hmp, iprec->leaf.data_offset, &error);
1828                         KKASSERT(error == 0);
1829                         hammer_del_buffers(hmp,
1830                                            iprec->leaf.data_offset,
1831                                            zone2_offset,
1832                                            iprec->leaf.data_len);
1833                 }
1834                 return(0);
1835         }
1836
1837         /*
1838          * On-disk records are marked as deleted by updating their delete_tid.
1839          * This does not effect their position in the B-Tree (which is based
1840          * on their create_tid).
1841          */
1842         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1843         elm = NULL;
1844
1845         /*
1846          * If we were mounted with the nohistory option, we physically
1847          * delete the record.
1848          */
1849         dodelete = hammer_nohistory(ip);
1850
1851         if (error == 0) {
1852                 error = hammer_cursor_upgrade(cursor);
1853                 if (error == 0) {
1854                         elm = &cursor->node->ondisk->elms[cursor->index];
1855                         hammer_modify_node(cursor->trans, cursor->node,
1856                                            elm, sizeof(*elm));
1857                         elm->leaf.base.delete_tid = tid;
1858                         elm->leaf.delete_ts = cursor->trans->time32;
1859                         hammer_modify_node_done(cursor->node);
1860
1861                         /*
1862                          * An on-disk record cannot have the same delete_tid
1863                          * as its create_tid.  In a chain of record updates
1864                          * this could result in a duplicate record.
1865                          */
1866                         KKASSERT(elm->leaf.base.delete_tid != elm->leaf.base.create_tid);
1867                 }
1868         }
1869
1870         if (error == 0 && dodelete) {
1871                 error = hammer_delete_at_cursor(cursor, NULL);
1872                 if (error) {
1873                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1874                         error = 0;
1875                 }
1876         }
1877         return(error);
1878 }
1879
1880 int
1881 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
1882 {
1883         hammer_btree_elm_t elm;
1884         hammer_off_t data_offset;
1885         int32_t data_len;
1886         u_int16_t rec_type;
1887         int error;
1888
1889         elm = &cursor->node->ondisk->elms[cursor->index];
1890         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1891
1892         data_offset = elm->leaf.data_offset;
1893         data_len = elm->leaf.data_len;
1894         rec_type = elm->leaf.base.rec_type;
1895
1896         error = hammer_btree_delete(cursor);
1897         if (error == 0) {
1898                 /*
1899                  * This forces a fixup for the iteration because
1900                  * the cursor is now either sitting at the 'next'
1901                  * element or sitting at the end of a leaf.
1902                  */
1903                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1904                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
1905                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1906                 }
1907         }
1908         if (error == 0) {
1909                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
1910                 case HAMMER_ZONE_LARGE_DATA:
1911                 case HAMMER_ZONE_SMALL_DATA:
1912                 case HAMMER_ZONE_META:
1913                         hammer_blockmap_free(cursor->trans,
1914                                              data_offset, data_len);
1915                         break;
1916                 default:
1917                         break;
1918                 }
1919         }
1920         return (error);
1921 }
1922
1923 /*
1924  * Determine whether we can remove a directory.  This routine checks whether
1925  * a directory is empty or not and enforces flush connectivity.
1926  *
1927  * Flush connectivity requires that we block if the target directory is
1928  * currently flushing, otherwise it may not end up in the same flush group.
1929  *
1930  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
1931  */
1932 int
1933 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
1934 {
1935         struct hammer_cursor cursor;
1936         int error;
1937
1938         /*
1939          * Check directory empty
1940          */
1941         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1942
1943         cursor.key_beg.localization = ip->obj_localization +
1944                                       HAMMER_LOCALIZE_MISC;
1945         cursor.key_beg.obj_id = ip->obj_id;
1946         cursor.key_beg.create_tid = 0;
1947         cursor.key_beg.delete_tid = 0;
1948         cursor.key_beg.obj_type = 0;
1949         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1950         cursor.key_beg.key = HAMMER_MIN_KEY;
1951
1952         cursor.key_end = cursor.key_beg;
1953         cursor.key_end.rec_type = 0xFFFF;
1954         cursor.key_end.key = HAMMER_MAX_KEY;
1955
1956         cursor.asof = ip->obj_asof;
1957         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1958
1959         error = hammer_ip_first(&cursor);
1960         if (error == ENOENT)
1961                 error = 0;
1962         else if (error == 0)
1963                 error = ENOTEMPTY;
1964         hammer_done_cursor(&cursor);
1965         return(error);
1966 }
1967