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 1.76 2008/06/26 04:06:23 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         struct hammer_cursor cursor;
 594         hammer_record_t record;
 595         int error;
 596         int count;
 597         u_int32_t iterator;
 598
 599         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 600         if (++trans->hmp->namekey_iterator == 0)
 601                 ++trans->hmp->namekey_iterator;
 602
 603         record->type = HAMMER_MEM_RECORD_ADD;
 604         record->leaf.base.localization = dip->obj_localization +
 605                                          HAMMER_LOCALIZE_MISC;
 606         record->leaf.base.obj_id = dip->obj_id;
 607         record->leaf.base.key = hammer_directory_namekey(name, bytes);
 608         record->leaf.base.key += trans->hmp->namekey_iterator;
 609         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 610         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 611         record->data->entry.obj_id = ip->obj_id;
 612         record->data->entry.localization = ip->obj_localization;
 613         bcopy(name, record->data->entry.name, bytes);
 614
 615         ++ip->ino_data.nlinks;
 616         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 617
 618         /*
 619          * Find an unused namekey.  Both the in-memory record tree and
 620          * the B-Tree are checked.  Exact matches also match create_tid
 621          * so use an ASOF search to (mostly) ignore it.
 622          */
 623         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 624         cursor.key_beg = record->leaf.base;
 625         cursor.flags |= HAMMER_CURSOR_ASOF;
 626         cursor.asof = ip->obj_asof;
 627
 628         count = 0;
 629         while (hammer_ip_lookup(&cursor) == 0) {
 630                 iterator = (u_int32_t)record->leaf.base.key + 1;
 631                 if (iterator == 0)
 632                         iterator = 1;
 633                 record->leaf.base.key &= ~0xFFFFFFFFLL;
 634                 record->leaf.base.key |= iterator;
 635                 cursor.key_beg.key = record->leaf.base.key;
 636                 if (++count == 1000000000) {
 637                         hammer_rel_mem_record(record);
 638                         error = ENOSPC;
 639                         goto failed;
 640                 }
 641         }
 642
 643         /*
 644          * The target inode and the directory entry are bound together.
 645          */
 646         record->target_ip = ip;
 647         record->flush_state = HAMMER_FST_SETUP;
 648         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 649
 650         /*
 651          * The inode now has a dependancy and must be taken out of the idle
 652          * state.  An inode not in an idle state is given an extra reference.
 653          */
 654         if (ip->flush_state == HAMMER_FST_IDLE) {
 655                 hammer_ref(&ip->lock);
 656                 ip->flush_state = HAMMER_FST_SETUP;
 657         }
 658         error = hammer_mem_add(record);
 659 failed:
 660         hammer_done_cursor(&cursor);
 661         return(error);
 662 }
 663
 664 /*
 665  * Delete the directory entry and update the inode link count.  The
 666  * cursor must be seeked to the directory entry record being deleted.
 667  *
 668  * The related inode should be share-locked by the caller.  The caller is
 669  * on the frontend.
 670  *
 671  * This function can return EDEADLK requiring the caller to terminate
 672  * the cursor, any locks, wait on the returned record, and retry.
 673  */
 674 int
 675 hammer_ip_del_directory(struct hammer_transaction *trans,
 676                      hammer_cursor_t cursor, struct hammer_inode *dip,
 677                      struct hammer_inode *ip)
 678 {
 679         hammer_record_t record;
 680         int error;
 681
 682         if (hammer_cursor_inmem(cursor)) {
 683                 /*
 684                  * In-memory (unsynchronized) records can simply be freed.
 685                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 686                  * by the backend, we must still avoid races against the
 687                  * backend potentially syncing the record to the media.
 688                  *
 689                  * We cannot call hammer_ip_delete_record(), that routine may
 690                  * only be called from the backend.
 691                  */
 692                 record = cursor->iprec;
 693                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 694                         KKASSERT(cursor->deadlk_rec == NULL);
 695                         hammer_ref(&record->lock);
 696                         cursor->deadlk_rec = record;
 697                         error = EDEADLK;
 698                 } else {
 699                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 700                         record->flags |= HAMMER_RECF_DELETED_FE;
 701                         error = 0;
 702                 }
 703         } else {
 704                 /*
 705                  * If the record is on-disk we have to queue the deletion by
 706                  * the record's key.  This also causes lookups to skip the
 707                  * record.
 708                  */
 709                 KKASSERT(dip->flags &
 710                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 711                 record = hammer_alloc_mem_record(dip, 0);
 712                 record->type = HAMMER_MEM_RECORD_DEL;
 713                 record->leaf.base = cursor->leaf->base;
 714
 715                 record->target_ip = ip;
 716                 record->flush_state = HAMMER_FST_SETUP;
 717                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 718
 719                 /*
 720                  * The inode now has a dependancy and must be taken out of
 721                  * the idle state.  An inode not in an idle state is given
 722                  * an extra reference.
 723                  */
 724                 if (ip->flush_state == HAMMER_FST_IDLE) {
 725                         hammer_ref(&ip->lock);
 726                         ip->flush_state = HAMMER_FST_SETUP;
 727                 }
 728
 729                 error = hammer_mem_add(record);
 730         }
 731
 732         /*
 733          * One less link.  The file may still be open in the OS even after
 734          * all links have gone away.
 735          *
 736          * We have to terminate the cursor before syncing the inode to
 737          * avoid deadlocking against ourselves.  XXX this may no longer
 738          * be true.
 739          *
 740          * If nlinks drops to zero and the vnode is inactive (or there is
 741          * no vnode), call hammer_inode_unloadable_check() to zonk the
 742          * inode.  If we don't do this here the inode will not be destroyed
 743          * on-media until we unmount.
 744          */
 745         if (error == 0) {
 746                 --ip->ino_data.nlinks;
 747                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 748                 if (ip->ino_data.nlinks == 0 &&
 749                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 750                         hammer_done_cursor(cursor);
 751                         hammer_inode_unloadable_check(ip, 1);
 752                         hammer_flush_inode(ip, 0);
 753                 }
 754
 755         }
 756         return(error);
 757 }
 758
 759 /*
 760  * Add a record to an inode.
 761  *
 762  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 763  * initialize the following additional fields:
 764  *
 765  * The related inode should be share-locked by the caller.  The caller is
 766  * on the frontend.
 767  *
 768  * record->rec.entry.base.base.key
 769  * record->rec.entry.base.base.rec_type
 770  * record->rec.entry.base.base.data_len
 771  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 772  */
 773 int
 774 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 775 {
 776         hammer_inode_t ip = record->ip;
 777         int error;
 778
 779         KKASSERT(record->leaf.base.localization != 0);
 780         record->leaf.base.obj_id = ip->obj_id;
 781         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 782         error = hammer_mem_add(record);
 783         return(error);
 784 }
 785
 786 /*
 787  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 788  * reserved so the front-end can flush large data writes without having
 789  * to queue the BIO to the flusher.  Only the related record gets queued
 790  * to the flusher.
 791  */
 792 static hammer_record_t
 793 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 794 {
 795         hammer_record_t record;
 796         struct hammer_btree_leaf_elm leaf;
 797
 798         bzero(&leaf, sizeof(leaf));
 799         leaf.base.obj_id = ip->obj_id;
 800         leaf.base.key = file_offset + bytes;
 801         leaf.base.create_tid = 0;
 802         leaf.base.delete_tid = 0;
 803         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 804         leaf.base.obj_type = 0;                 /* unused */
 805         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
 806         leaf.base.localization = ip->obj_localization + HAMMER_LOCALIZE_MISC;
 807         leaf.data_len = bytes;
 808
 809         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
 810         if (record)
 811                 hammer_ref(&record->lock);
 812         return(record);
 813 }
 814
 815 /*
 816  * Reserve blockmap space placemarked with an in-memory record.
 817  *
 818  * This routine is called by the frontend in order to be able to directly
 819  * flush a buffer cache buffer.  The frontend has locked the related buffer
 820  * cache buffers and we should be able to manipulate any overlapping
 821  * in-memory records.
 822  */
 823 hammer_record_t
 824 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 825                    int *errorp)
 826 {
 827         hammer_record_t record;
 828         hammer_record_t conflict;
 829         int zone;
 830         int flags;
 831
 832         /*
 833          * Deal with conflicting in-memory records.  We cannot have multiple
 834          * in-memory records for the same offset without seriously confusing
 835          * the backend, including but not limited to the backend issuing
 836          * delete-create-delete sequences and asserting on the delete_tid
 837          * being the same as the create_tid.
 838          *
 839          * If we encounter a record with the backend interlock set we cannot
 840          * immediately delete it without confusing the backend.
 841          */
 842         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 843                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 844                         conflict->flags |= HAMMER_RECF_WANTED;
 845                         tsleep(conflict, 0, "hmrrc3", 0);
 846                 } else {
 847                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 848                 }
 849                 hammer_rel_mem_record(conflict);
 850         }
 851
 852         /*
 853          * Create a record to cover the direct write.  This is called with
 854          * the related BIO locked so there should be no possible conflict.
 855          *
 856          * The backend is responsible for finalizing the space reserved in
 857          * this record.
 858          *
 859          * XXX bytes not aligned, depend on the reservation code to
 860          * align the reservation.
 861          */
 862         record = hammer_alloc_mem_record(ip, 0);
 863         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 864                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 865         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 866                                                &record->leaf.data_offset,
 867                                                errorp);
 868         if (record->resv == NULL) {
 869                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 870                 hammer_rel_mem_record(record);
 871                 return(NULL);
 872         }
 873         record->type = HAMMER_MEM_RECORD_DATA;
 874         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 875         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 876         record->leaf.base.obj_id = ip->obj_id;
 877         record->leaf.base.key = file_offset + bytes;
 878         record->leaf.base.localization = ip->obj_localization +
 879                                          HAMMER_LOCALIZE_MISC;
 880         record->leaf.data_len = bytes;
 881         hammer_crc_set_leaf(data, &record->leaf);
 882         flags = record->flags;
 883
 884         hammer_ref(&record->lock);      /* mem_add eats a reference */
 885         *errorp = hammer_mem_add(record);
 886         if (*errorp) {
 887                 conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 888                 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
 889                         *errorp, conflict, file_offset, bytes);
 890                 if (conflict)
 891                         kprintf("conflict %lld %d\n", conflict->leaf.base.key, conflict->leaf.data_len);
 892                 if (conflict)
 893                         hammer_rel_mem_record(conflict);
 894         }
 895         KKASSERT(*errorp == 0);
 896         conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 897         if (conflict != record) {
 898                 kprintf("conflict mismatch %p %p %08x\n", conflict, record, record->flags);
 899                 if (conflict)
 900                     kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict->leaf.base.key, conflict->leaf.data_len, record->leaf.base.key, record->leaf.data_len);
 901         }
 902         KKASSERT(conflict == record);
 903         hammer_rel_mem_record(conflict);
 904
 905         return (record);
 906 }
 907
 908 /*
 909  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 910  * and records in a flushing state are handled by the backend.  The vnops
 911  * setattr code will handle the block containing the truncation point.
 912  *
 913  * Partial blocks are not deleted.
 914  */
 915 int
 916 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 917 {
 918         struct rec_trunc_info info;
 919
 920         switch(ip->ino_data.obj_type) {
 921         case HAMMER_OBJTYPE_REGFILE:
 922                 info.rec_type = HAMMER_RECTYPE_DATA;
 923                 break;
 924         case HAMMER_OBJTYPE_DBFILE:
 925                 info.rec_type = HAMMER_RECTYPE_DB;
 926                 break;
 927         default:
 928                 return(EINVAL);
 929         }
 930         info.trunc_off = file_size;
 931         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 932                                    hammer_rec_trunc_callback, &info);
 933         return(0);
 934 }
 935
 936 static int
 937 hammer_rec_trunc_callback(hammer_record_t record, void *data __unused)
 938 {
 939         if (record->flags & HAMMER_RECF_DELETED_FE)
 940                 return(0);
 941         if (record->flush_state == HAMMER_FST_FLUSH)
 942                 return(0);
 943         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 944         hammer_ref(&record->lock);
 945         record->flags |= HAMMER_RECF_DELETED_FE;
 946         hammer_rel_mem_record(record);
 947         return(0);
 948 }
 949
 950 /*
 951  * Return 1 if the caller must check for and delete existing records
 952  * before writing out a new data record.
 953  *
 954  * Return 0 if the caller can just insert the record into the B-Tree without
 955  * checking.
 956  */
 957 static int
 958 hammer_record_needs_overwrite_delete(hammer_record_t record)
 959 {
 960         hammer_inode_t ip = record->ip;
 961         int64_t file_offset;
 962         int r;
 963
 964         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
 965                 file_offset = record->leaf.base.key;
 966         else
 967                 file_offset = record->leaf.base.key - record->leaf.data_len;
 968         r = (file_offset < ip->sync_trunc_off);
 969         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
 970                 if (ip->sync_trunc_off <= record->leaf.base.key)
 971                         ip->sync_trunc_off = record->leaf.base.key + 1;
 972         } else {
 973                 if (ip->sync_trunc_off < record->leaf.base.key)
 974                         ip->sync_trunc_off = record->leaf.base.key;
 975         }
 976         return(r);
 977 }
 978
 979 /*
 980  * Backend code.  Sync a record to the media.
 981  */
 982 int
 983 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
 984 {
 985         hammer_transaction_t trans = cursor->trans;
 986         int64_t file_offset;
 987         int bytes;
 988         void *bdata;
 989         int error;
 990
 991         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 992         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 993         KKASSERT(record->leaf.base.localization != 0);
 994
 995         /*
 996          * If this is a bulk-data record placemarker there may be an existing
 997          * record on-disk, indicating a data overwrite.  If there is the
 998          * on-disk record must be deleted before we can insert our new record.
 999          *
1000          * We've synthesized this record and do not know what the create_tid
1001          * on-disk is, nor how much data it represents.
1002          *
1003          * Keep in mind that (key) for data records is (base_offset + len),
1004          * not (base_offset).  Also, we only want to get rid of on-disk
1005          * records since we are trying to sync our in-memory record, call
1006          * hammer_ip_delete_range() with truncating set to 1 to make sure
1007          * it skips in-memory records.
1008          *
1009          * It is ok for the lookup to return ENOENT.
1010          *
1011          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1012          * to call hammer_ip_delete_range() or not.  This also means we must
1013          * update sync_trunc_off() as we write.
1014          */
1015         if (record->type == HAMMER_MEM_RECORD_DATA &&
1016             hammer_record_needs_overwrite_delete(record)) {
1017                 file_offset = record->leaf.base.key - record->leaf.data_len;
1018                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1019                         ~HAMMER_BUFMASK;
1020                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1021                 error = hammer_ip_delete_range(
1022                                 cursor, record->ip,
1023                                 file_offset, file_offset + bytes - 1,
1024                                 1);
1025                 if (error && error != ENOENT)
1026                         goto done;
1027         }
1028
1029         /*
1030          * Setup the cursor.
1031          */
1032         hammer_normalize_cursor(cursor);
1033         cursor->key_beg = record->leaf.base;
1034         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1035         cursor->flags |= HAMMER_CURSOR_BACKEND;
1036         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1037
1038         /*
1039          * Records can wind up on-media before the inode itself is on-media.
1040          * Flag the case.
1041          */
1042         record->ip->flags |= HAMMER_INODE_DONDISK;
1043
1044         /*
1045          * If we are deleting a directory entry an exact match must be
1046          * found on-disk.
1047          */
1048         if (record->type == HAMMER_MEM_RECORD_DEL) {
1049                 error = hammer_btree_lookup(cursor);
1050                 if (error == 0) {
1051                         error = hammer_ip_delete_record(cursor, record->ip,
1052                                                         trans->tid);
1053                         if (error == 0) {
1054                                 record->flags |= HAMMER_RECF_DELETED_FE;
1055                                 record->flags |= HAMMER_RECF_DELETED_BE;
1056                         }
1057                 }
1058                 goto done;
1059         }
1060
1061         /*
1062          * We are inserting.
1063          *
1064          * Issue a lookup to position the cursor and locate the cluster.  The
1065          * target key should not exist.  If we are creating a directory entry
1066          * we may have to iterate the low 32 bits of the key to find an unused
1067          * key.
1068          */
1069         cursor->flags |= HAMMER_CURSOR_INSERT;
1070
1071         error = hammer_btree_lookup(cursor);
1072         if (hammer_debug_inode)
1073                 kprintf("DOINSERT LOOKUP %d\n", error);
1074         if (error == 0) {
1075                 kprintf("hammer_ip_sync_record: duplicate rec "
1076                         "at (%016llx)\n", record->leaf.base.key);
1077                 Debugger("duplicate record1");
1078                 error = EIO;
1079         }
1080 #if 0
1081         if (record->type == HAMMER_MEM_RECORD_DATA)
1082                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1083                         record->leaf.base.key - record->leaf.data_len,
1084                         record->leaf.data_offset, error);
1085 #endif
1086
1087         if (error != ENOENT)
1088                 goto done;
1089
1090         /*
1091          * Allocate the record and data.  The result buffers will be
1092          * marked as being modified and further calls to
1093          * hammer_modify_buffer() will result in unneeded UNDO records.
1094          *
1095          * Support zero-fill records (data == NULL and data_len != 0)
1096          */
1097         if (record->type == HAMMER_MEM_RECORD_DATA) {
1098                 /*
1099                  * The data portion of a bulk-data record has already been
1100                  * committed to disk, we need only adjust the layer2
1101                  * statistics in the same transaction as our B-Tree insert.
1102                  */
1103                 KKASSERT(record->leaf.data_offset != 0);
1104                 hammer_blockmap_finalize(trans, record->leaf.data_offset,
1105                                          record->leaf.data_len);
1106                 error = 0;
1107         } else if (record->data && record->leaf.data_len) {
1108                 /*
1109                  * Wholely cached record, with data.  Allocate the data.
1110                  */
1111                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1112                                           record->leaf.base.rec_type,
1113                                           &record->leaf.data_offset,
1114                                           &cursor->data_buffer, &error);
1115                 if (bdata == NULL)
1116                         goto done;
1117                 hammer_crc_set_leaf(record->data, &record->leaf);
1118                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1119                 bcopy(record->data, bdata, record->leaf.data_len);
1120                 hammer_modify_buffer_done(cursor->data_buffer);
1121         } else {
1122                 /*
1123                  * Wholely cached record, without data.
1124                  */
1125                 record->leaf.data_offset = 0;
1126                 record->leaf.data_crc = 0;
1127         }
1128
1129         error = hammer_btree_insert(cursor, &record->leaf);
1130         if (hammer_debug_inode && error)
1131                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1132
1133         /*
1134          * Our record is on-disk, normally mark the in-memory version as
1135          * deleted.  If the record represented a directory deletion but
1136          * we had to sync a valid directory entry to disk we must convert
1137          * the record to a covering delete so the frontend does not have
1138          * visibility on the synced entry.
1139          */
1140         if (error == 0) {
1141                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1142                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1143                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1144                         record->type = HAMMER_MEM_RECORD_DEL;
1145                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1146                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1147                         /* hammer_flush_record_done takes care of the rest */
1148                 } else {
1149                         record->flags |= HAMMER_RECF_DELETED_FE;
1150                         record->flags |= HAMMER_RECF_DELETED_BE;
1151                 }
1152         } else {
1153                 if (record->leaf.data_offset) {
1154                         hammer_blockmap_free(trans, record->leaf.data_offset,
1155                                              record->leaf.data_len);
1156                 }
1157         }
1158
1159 done:
1160         return(error);
1161 }
1162
1163 /*
1164  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1165  * entry's key is used to deal with hash collisions in the upper 32 bits.
1166  * A unique 64 bit key is generated in-memory and may be regenerated a
1167  * second time when the directory record is flushed to the on-disk B-Tree.
1168  *
1169  * A referenced record is passed to this function.  This function
1170  * eats the reference.  If an error occurs the record will be deleted.
1171  *
1172  * A copy of the temporary record->data pointer provided by the caller
1173  * will be made.
1174  */
1175 static
1176 int
1177 hammer_mem_add(hammer_record_t record)
1178 {
1179         hammer_mount_t hmp = record->ip->hmp;
1180
1181         /*
1182          * Make a private copy of record->data
1183          */
1184         if (record->data)
1185                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1186
1187         /*
1188          * Insert into the RB tree.  A unique key should have already
1189          * been selected if this is a directory entry.
1190          */
1191         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1192                 record->flags |= HAMMER_RECF_DELETED_FE;
1193                 hammer_rel_mem_record(record);
1194                 return (EEXIST);
1195         }
1196         ++hmp->count_newrecords;
1197         ++hmp->rsv_recs;
1198         ++record->ip->rsv_recs;
1199         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1200         record->flags |= HAMMER_RECF_ONRBTREE;
1201         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1202         hammer_rel_mem_record(record);
1203         return(0);
1204 }
1205
1206 /************************************************************************
1207  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1208  ************************************************************************
1209  *
1210  * These functions augment the B-Tree scanning functions in hammer_btree.c
1211  * by merging in-memory records with on-disk records.
1212  */
1213
1214 /*
1215  * Locate a particular record either in-memory or on-disk.
1216  *
1217  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1218  * NOT be called to iterate results.
1219  */
1220 int
1221 hammer_ip_lookup(hammer_cursor_t cursor)
1222 {
1223         int error;
1224
1225         /*
1226          * If the element is in-memory return it without searching the
1227          * on-disk B-Tree
1228          */
1229         KKASSERT(cursor->ip);
1230         error = hammer_mem_lookup(cursor);
1231         if (error == 0) {
1232                 cursor->leaf = &cursor->iprec->leaf;
1233                 return(error);
1234         }
1235         if (error != ENOENT)
1236                 return(error);
1237
1238         /*
1239          * If the inode has on-disk components search the on-disk B-Tree.
1240          */
1241         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1242                 return(error);
1243         error = hammer_btree_lookup(cursor);
1244         if (error == 0)
1245                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1246         return(error);
1247 }
1248
1249 /*
1250  * Locate the first record within the cursor's key_beg/key_end range,
1251  * restricted to a particular inode.  0 is returned on success, ENOENT
1252  * if no records matched the requested range, or some other error.
1253  *
1254  * When 0 is returned hammer_ip_next() may be used to iterate additional
1255  * records within the requested range.
1256  *
1257  * This function can return EDEADLK, requiring the caller to terminate
1258  * the cursor and try again.
1259  */
1260 int
1261 hammer_ip_first(hammer_cursor_t cursor)
1262 {
1263         hammer_inode_t ip = cursor->ip;
1264         int error;
1265
1266         KKASSERT(ip != NULL);
1267
1268         /*
1269          * Clean up fields and setup for merged scan
1270          */
1271         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1272         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1273         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1274         if (cursor->iprec) {
1275                 hammer_rel_mem_record(cursor->iprec);
1276                 cursor->iprec = NULL;
1277         }
1278
1279         /*
1280          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1281          * exact lookup so if we get ENOENT we have to call the iterate
1282          * function to validate the first record after the begin key.
1283          *
1284          * The ATEDISK flag is used by hammer_btree_iterate to determine
1285          * whether it must index forwards or not.  It is also used here
1286          * to select the next record from in-memory or on-disk.
1287          *
1288          * EDEADLK can only occur if the lookup hit an empty internal
1289          * element and couldn't delete it.  Since this could only occur
1290          * in-range, we can just iterate from the failure point.
1291          */
1292         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1293                 error = hammer_btree_lookup(cursor);
1294                 if (error == ENOENT || error == EDEADLK) {
1295                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1296                         if (hammer_debug_general & 0x2000)
1297                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1298                         error = hammer_btree_iterate(cursor);
1299                 }
1300                 if (error && error != ENOENT)
1301                         return(error);
1302                 if (error == 0) {
1303                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1304                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1305                 } else {
1306                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1307                 }
1308         }
1309
1310         /*
1311          * Search the in-memory record list (Red-Black tree).  Unlike the
1312          * B-Tree search, mem_first checks for records in the range.
1313          */
1314         error = hammer_mem_first(cursor);
1315         if (error && error != ENOENT)
1316                 return(error);
1317         if (error == 0) {
1318                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1319                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1320                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1321                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1322         }
1323
1324         /*
1325          * This will return the first matching record.
1326          */
1327         return(hammer_ip_next(cursor));
1328 }
1329
1330 /*
1331  * Retrieve the next record in a merged iteration within the bounds of the
1332  * cursor.  This call may be made multiple times after the cursor has been
1333  * initially searched with hammer_ip_first().
1334  *
1335  * 0 is returned on success, ENOENT if no further records match the
1336  * requested range, or some other error code is returned.
1337  */
1338 int
1339 hammer_ip_next(hammer_cursor_t cursor)
1340 {
1341         hammer_btree_elm_t elm;
1342         hammer_record_t rec, save;
1343         int error;
1344         int r;
1345
1346 next_btree:
1347         /*
1348          * Load the current on-disk and in-memory record.  If we ate any
1349          * records we have to get the next one.
1350          *
1351          * If we deleted the last on-disk record we had scanned ATEDISK will
1352          * be clear and DELBTREE will be set, forcing a call to iterate. The
1353          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1354          * element.  If ATEDISK is set, iterate will skip the 'current'
1355          * element.
1356          *
1357          * Get the next on-disk record
1358          */
1359         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1360                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1361                         error = hammer_btree_iterate(cursor);
1362                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1363                         if (error == 0) {
1364                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1365                                 hammer_cache_node(&cursor->ip->cache[1],
1366                                                   cursor->node);
1367                         } else {
1368                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1369                                                  HAMMER_CURSOR_ATEDISK;
1370                         }
1371                 }
1372         }
1373
1374 next_memory:
1375         /*
1376          * Get the next in-memory record.  The record can be ripped out
1377          * of the RB tree so we maintain a scan_info structure to track
1378          * the next node.
1379          *
1380          * hammer_rec_scan_cmp:  Is the record still in our general range,
1381          *                       (non-inclusive of snapshot exclusions)?
1382          * hammer_rec_scan_callback: Is the record in our snapshot?
1383          */
1384         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1385                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1386                         save = cursor->iprec;
1387                         cursor->iprec = NULL;
1388                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1389                         while (rec) {
1390                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1391                                         break;
1392                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1393                                         break;
1394                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1395                         }
1396                         if (save)
1397                                 hammer_rel_mem_record(save);
1398                         if (cursor->iprec) {
1399                                 KKASSERT(cursor->iprec == rec);
1400                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1401                         } else {
1402                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1403                         }
1404                 }
1405         }
1406
1407         /*
1408          * The memory record may have become stale while being held in
1409          * cursor->iprec.  We are interlocked against the backend on
1410          * with regards to B-Tree entries.
1411          */
1412         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1413                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1414                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1415                         goto next_memory;
1416                 }
1417         }
1418
1419         /*
1420          * Extract either the disk or memory record depending on their
1421          * relative position.
1422          */
1423         error = 0;
1424         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1425         case 0:
1426                 /*
1427                  * Both entries valid.   Compare the entries and nominally
1428                  * return the first one in the sort order.  Numerous cases
1429                  * require special attention, however.
1430                  */
1431                 elm = &cursor->node->ondisk->elms[cursor->index];
1432                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1433
1434                 /*
1435                  * If the two entries differ only by their key (-2/2) or
1436                  * create_tid (-1/1), and are DATA records, we may have a
1437                  * nominal match.  We have to calculate the base file
1438                  * offset of the data.
1439                  */
1440                 if (r <= 2 && r >= -2 && r != 0 &&
1441                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1442                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1443                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1444                         int64_t base2 = cursor->iprec->leaf.base.key -
1445                                         cursor->iprec->leaf.data_len;
1446                         if (base1 == base2)
1447                                 r = 0;
1448                 }
1449
1450                 if (r < 0) {
1451                         error = hammer_btree_extract(cursor,
1452                                                      HAMMER_CURSOR_GET_LEAF);
1453                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1454                         break;
1455                 }
1456
1457                 /*
1458                  * If the entries match exactly the memory entry is either
1459                  * an on-disk directory entry deletion or a bulk data
1460                  * overwrite.  If it is a directory entry deletion we eat
1461                  * both entries.
1462                  *
1463                  * For the bulk-data overwrite case it is possible to have
1464                  * visibility into both, which simply means the syncer
1465                  * hasn't gotten around to doing the delete+insert sequence
1466                  * on the B-Tree.  Use the memory entry and throw away the
1467                  * on-disk entry.
1468                  *
1469                  * If the in-memory record is not either of these we
1470                  * probably caught the syncer while it was syncing it to
1471                  * the media.  Since we hold a shared lock on the cursor,
1472                  * the in-memory record had better be marked deleted at
1473                  * this point.
1474                  */
1475                 if (r == 0) {
1476                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1477                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1478                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1479                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1480                                         goto next_btree;
1481                                 }
1482                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1483                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1484                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1485                                 }
1486                                 /* fall through to memory entry */
1487                         } else {
1488                                 panic("hammer_ip_next: duplicate mem/b-tree entry");
1489                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1490                                 goto next_memory;
1491                         }
1492                 }
1493                 /* fall through to the memory entry */
1494         case HAMMER_CURSOR_ATEDISK:
1495                 /*
1496                  * Only the memory entry is valid.
1497                  */
1498                 cursor->leaf = &cursor->iprec->leaf;
1499                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1500
1501                 /*
1502                  * If the memory entry is an on-disk deletion we should have
1503                  * also had found a B-Tree record.  If the backend beat us
1504                  * to it it would have interlocked the cursor and we should
1505                  * have seen the in-memory record marked DELETED_FE.
1506                  */
1507                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1508                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1509                         panic("hammer_ip_next: del-on-disk with no b-tree entry");
1510                 }
1511                 break;
1512         case HAMMER_CURSOR_ATEMEM:
1513                 /*
1514                  * Only the disk entry is valid
1515                  */
1516                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1517                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1518                 break;
1519         default:
1520                 /*
1521                  * Neither entry is valid
1522                  *
1523                  * XXX error not set properly
1524                  */
1525                 cursor->leaf = NULL;
1526                 error = ENOENT;
1527                 break;
1528         }
1529         return(error);
1530 }
1531
1532 /*
1533  * Resolve the cursor->data pointer for the current cursor position in
1534  * a merged iteration.
1535  */
1536 int
1537 hammer_ip_resolve_data(hammer_cursor_t cursor)
1538 {
1539         hammer_record_t record;
1540         int error;
1541
1542         if (hammer_cursor_inmem(cursor)) {
1543                 /*
1544                  * The data associated with an in-memory record is usually
1545                  * kmalloced, but reserve-ahead data records will have an
1546                  * on-disk reference.
1547                  *
1548                  * NOTE: Reserve-ahead data records must be handled in the
1549                  * context of the related high level buffer cache buffer
1550                  * to interlock against async writes.
1551                  */
1552                 record = cursor->iprec;
1553                 cursor->data = record->data;
1554                 error = 0;
1555                 if (cursor->data == NULL) {
1556                         KKASSERT(record->leaf.base.rec_type ==
1557                                  HAMMER_RECTYPE_DATA);
1558                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1559                                                     record->leaf.data_offset,
1560                                                     record->leaf.data_len,
1561                                                     &error,
1562                                                     &cursor->data_buffer);
1563                 }
1564         } else {
1565                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1566                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1567         }
1568         return(error);
1569 }
1570
1571 /*
1572  * Backend truncation / record replacement - delete records in range.
1573  *
1574  * Delete all records within the specified range for inode ip.  In-memory
1575  * records still associated with the frontend are ignored.
1576  *
1577  * NOTE: An unaligned range will cause new records to be added to cover
1578  * the edge cases. (XXX not implemented yet).
1579  *
1580  * NOTE: Replacement via reservations (see hammer_ip_sync_record_cursor())
1581  * also do not deal with unaligned ranges.
1582  *
1583  * NOTE: ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1584  *
1585  * NOTE: Record keys for regular file data have to be special-cased since
1586  * they indicate the end of the range (key = base + bytes).
1587  */
1588 int
1589 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1590                        int64_t ran_beg, int64_t ran_end, int truncating)
1591 {
1592         hammer_transaction_t trans = cursor->trans;
1593         hammer_btree_leaf_elm_t leaf;
1594         int error;
1595         int64_t off;
1596
1597 #if 0
1598         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1599 #endif
1600
1601         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1602 retry:
1603         hammer_normalize_cursor(cursor);
1604         cursor->key_beg.localization = ip->obj_localization +
1605                                        HAMMER_LOCALIZE_MISC;
1606         cursor->key_beg.obj_id = ip->obj_id;
1607         cursor->key_beg.create_tid = 0;
1608         cursor->key_beg.delete_tid = 0;
1609         cursor->key_beg.obj_type = 0;
1610         cursor->asof = ip->obj_asof;
1611         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1612         cursor->flags |= HAMMER_CURSOR_ASOF;
1613         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1614         cursor->flags |= HAMMER_CURSOR_BACKEND;
1615
1616         cursor->key_end = cursor->key_beg;
1617         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1618                 cursor->key_beg.key = ran_beg;
1619                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1620                 cursor->key_end.rec_type = HAMMER_RECTYPE_DB;
1621                 cursor->key_end.key = ran_end;
1622         } else {
1623                 /*
1624                  * The key in the B-Tree is (base+bytes), so the first possible
1625                  * matching key is ran_beg + 1.
1626                  */
1627                 int64_t tmp64;
1628
1629                 cursor->key_beg.key = ran_beg + 1;
1630                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1631                 cursor->key_end.rec_type = HAMMER_RECTYPE_DATA;
1632
1633                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1634                 if (tmp64 < ran_end)
1635                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1636                 else
1637                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1638         }
1639         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1640
1641         error = hammer_ip_first(cursor);
1642
1643         /*
1644          * Iterate through matching records and mark them as deleted.
1645          */
1646         while (error == 0) {
1647                 leaf = cursor->leaf;
1648
1649                 KKASSERT(leaf->base.delete_tid == 0);
1650
1651                 /*
1652                  * There may be overlap cases for regular file data.  Also
1653                  * remember the key for a regular file record is (base + len),
1654                  * NOT (base).
1655                  */
1656                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1657                         off = leaf->base.key - leaf->data_len;
1658                         /*
1659                          * Check the left edge case.  We currently do not
1660                          * split existing records.
1661                          */
1662                         if (off < ran_beg) {
1663                                 panic("hammer left edge case %016llx %d\n",
1664                                         leaf->base.key, leaf->data_len);
1665                         }
1666
1667                         /*
1668                          * Check the right edge case.  Note that the
1669                          * record can be completely out of bounds, which
1670                          * terminates the search.
1671                          *
1672                          * base->key is exclusive of the right edge while
1673                          * ran_end is inclusive of the right edge.  The
1674                          * (key - data_len) left boundary is inclusive.
1675                          *
1676                          * XXX theory-check this test at some point, are
1677                          * we missing a + 1 somewhere?  Note that ran_end
1678                          * could overflow.
1679                          */
1680                         if (leaf->base.key - 1 > ran_end) {
1681                                 if (leaf->base.key - leaf->data_len > ran_end)
1682                                         break;
1683                                 panic("hammer right edge case\n");
1684                         }
1685                 }
1686
1687                 /*
1688                  * Delete the record.  When truncating we do not delete
1689                  * in-memory (data) records because they represent data
1690                  * written after the truncation.
1691                  *
1692                  * This will also physically destroy the B-Tree entry and
1693                  * data if the retention policy dictates.  The function
1694                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1695                  * uses to perform a fixup.
1696                  */
1697                 if (truncating == 0 || hammer_cursor_ondisk(cursor))
1698                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1699                 if (error)
1700                         break;
1701                 error = hammer_ip_next(cursor);
1702         }
1703         if (cursor->node)
1704                 hammer_cache_node(&ip->cache[1], cursor->node);
1705
1706         if (error == EDEADLK) {
1707                 hammer_done_cursor(cursor);
1708                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1709                 if (error == 0)
1710                         goto retry;
1711         }
1712         if (error == ENOENT)
1713                 error = 0;
1714         return(error);
1715 }
1716
1717 /*
1718  * Backend truncation - delete all records.
1719  *
1720  * Delete all user records associated with an inode except the inode record
1721  * itself.  Directory entries are not deleted (they must be properly disposed
1722  * of or nlinks would get upset).
1723  */
1724 int
1725 hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
1726                            int *countp)
1727 {
1728         hammer_transaction_t trans = cursor->trans;
1729         hammer_btree_leaf_elm_t leaf;
1730         int error;
1731
1732         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1733 retry:
1734         hammer_normalize_cursor(cursor);
1735         cursor->key_beg.localization = ip->obj_localization +
1736                                        HAMMER_LOCALIZE_MISC;
1737         cursor->key_beg.obj_id = ip->obj_id;
1738         cursor->key_beg.create_tid = 0;
1739         cursor->key_beg.delete_tid = 0;
1740         cursor->key_beg.obj_type = 0;
1741         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1742         cursor->key_beg.key = HAMMER_MIN_KEY;
1743
1744         cursor->key_end = cursor->key_beg;
1745         cursor->key_end.rec_type = 0xFFFF;
1746         cursor->key_end.key = HAMMER_MAX_KEY;
1747
1748         cursor->asof = ip->obj_asof;
1749         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1750         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1751         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1752         cursor->flags |= HAMMER_CURSOR_BACKEND;
1753
1754         error = hammer_ip_first(cursor);
1755
1756         /*
1757          * Iterate through matching records and mark them as deleted.
1758          */
1759         while (error == 0) {
1760                 leaf = cursor->leaf;
1761
1762                 KKASSERT(leaf->base.delete_tid == 0);
1763
1764                 /*
1765                  * Mark the record and B-Tree entry as deleted.  This will
1766                  * also physically delete the B-Tree entry, record, and
1767                  * data if the retention policy dictates.  The function
1768                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1769                  * uses to perform a fixup.
1770                  *
1771                  * Directory entries (and delete-on-disk directory entries)
1772                  * must be synced and cannot be deleted.
1773                  */
1774                 if (leaf->base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
1775                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1776                         ++*countp;
1777                 }
1778                 if (error)
1779                         break;
1780                 error = hammer_ip_next(cursor);
1781         }
1782         if (cursor->node)
1783                 hammer_cache_node(&ip->cache[1], cursor->node);
1784         if (error == EDEADLK) {
1785                 hammer_done_cursor(cursor);
1786                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1787                 if (error == 0)
1788                         goto retry;
1789         }
1790         if (error == ENOENT)
1791                 error = 0;
1792         return(error);
1793 }
1794
1795 /*
1796  * Delete the record at the current cursor.  On success the cursor will
1797  * be positioned appropriately for an iteration but may no longer be at
1798  * a leaf node.
1799  *
1800  * This routine is only called from the backend.
1801  *
1802  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1803  * cursor and retry.
1804  */
1805 int
1806 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1807                         hammer_tid_t tid)
1808 {
1809         hammer_off_t zone2_offset;
1810         hammer_record_t iprec;
1811         hammer_btree_elm_t elm;
1812         hammer_mount_t hmp;
1813         int error;
1814         int dodelete;
1815
1816         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1817         KKASSERT(tid != 0);
1818         hmp = cursor->node->hmp;
1819
1820         /*
1821          * In-memory (unsynchronized) records can simply be freed.  This
1822          * only occurs in range iterations since all other records are
1823          * individually synchronized.  Thus there should be no confusion with
1824          * the interlock.
1825          *
1826          * An in-memory record may be deleted before being committed to disk,
1827          * but could have been accessed in the mean time.  The backing store
1828          * may never been marked allocated and so hammer_blockmap_free() may
1829          * never get called on it.  Because of this we have to make sure that
1830          * we've gotten rid of any related hammer_buffer or buffer cache
1831          * buffer.
1832          */
1833         if (hammer_cursor_inmem(cursor)) {
1834                 iprec = cursor->iprec;
1835                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1836                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1837                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1838
1839                 if (iprec->leaf.data_offset && iprec->leaf.data_len) {
1840                         zone2_offset = hammer_blockmap_lookup(hmp, iprec->leaf.data_offset, &error);
1841                         KKASSERT(error == 0);
1842                         hammer_del_buffers(hmp,
1843                                            iprec->leaf.data_offset,
1844                                            zone2_offset,
1845                                            iprec->leaf.data_len);
1846                 }
1847                 return(0);
1848         }
1849
1850         /*
1851          * On-disk records are marked as deleted by updating their delete_tid.
1852          * This does not effect their position in the B-Tree (which is based
1853          * on their create_tid).
1854          */
1855         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1856         elm = NULL;
1857
1858         /*
1859          * If we were mounted with the nohistory option, we physically
1860          * delete the record.
1861          */
1862         dodelete = hammer_nohistory(ip);
1863
1864         if (error == 0) {
1865                 error = hammer_cursor_upgrade(cursor);
1866                 if (error == 0) {
1867                         elm = &cursor->node->ondisk->elms[cursor->index];
1868                         hammer_modify_node(cursor->trans, cursor->node,
1869                                            elm, sizeof(*elm));
1870                         elm->leaf.base.delete_tid = tid;
1871                         elm->leaf.delete_ts = cursor->trans->time32;
1872                         hammer_modify_node_done(cursor->node);
1873
1874                         /*
1875                          * An on-disk record cannot have the same delete_tid
1876                          * as its create_tid.  In a chain of record updates
1877                          * this could result in a duplicate record.
1878                          */
1879                         KKASSERT(elm->leaf.base.delete_tid != elm->leaf.base.create_tid);
1880                 }
1881         }
1882
1883         if (error == 0 && dodelete) {
1884                 error = hammer_delete_at_cursor(cursor, NULL);
1885                 if (error) {
1886                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1887                         error = 0;
1888                 }
1889         }
1890         return(error);
1891 }
1892
1893 int
1894 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
1895 {
1896         hammer_btree_elm_t elm;
1897         hammer_off_t data_offset;
1898         int32_t data_len;
1899         u_int16_t rec_type;
1900         int error;
1901
1902         elm = &cursor->node->ondisk->elms[cursor->index];
1903         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1904
1905         data_offset = elm->leaf.data_offset;
1906         data_len = elm->leaf.data_len;
1907         rec_type = elm->leaf.base.rec_type;
1908
1909         error = hammer_btree_delete(cursor);
1910         if (error == 0) {
1911                 /*
1912                  * This forces a fixup for the iteration because
1913                  * the cursor is now either sitting at the 'next'
1914                  * element or sitting at the end of a leaf.
1915                  */
1916                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1917                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
1918                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1919                 }
1920         }
1921         if (error == 0) {
1922                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
1923                 case HAMMER_ZONE_LARGE_DATA:
1924                 case HAMMER_ZONE_SMALL_DATA:
1925                 case HAMMER_ZONE_META:
1926                         hammer_blockmap_free(cursor->trans,
1927                                              data_offset, data_len);
1928                         break;
1929                 default:
1930                         break;
1931                 }
1932         }
1933         return (error);
1934 }
1935
1936 /*
1937  * Determine whether we can remove a directory.  This routine checks whether
1938  * a directory is empty or not and enforces flush connectivity.
1939  *
1940  * Flush connectivity requires that we block if the target directory is
1941  * currently flushing, otherwise it may not end up in the same flush group.
1942  *
1943  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
1944  */
1945 int
1946 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
1947 {
1948         struct hammer_cursor cursor;
1949         int error;
1950
1951         /*
1952          * Check directory empty
1953          */
1954         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1955
1956         cursor.key_beg.localization = ip->obj_localization +
1957                                       HAMMER_LOCALIZE_MISC;
1958         cursor.key_beg.obj_id = ip->obj_id;
1959         cursor.key_beg.create_tid = 0;
1960         cursor.key_beg.delete_tid = 0;
1961         cursor.key_beg.obj_type = 0;
1962         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1963         cursor.key_beg.key = HAMMER_MIN_KEY;
1964
1965         cursor.key_end = cursor.key_beg;
1966         cursor.key_end.rec_type = 0xFFFF;
1967         cursor.key_end.key = HAMMER_MAX_KEY;
1968
1969         cursor.asof = ip->obj_asof;
1970         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1971
1972         error = hammer_ip_first(&cursor);
1973         if (error == ENOENT)
1974                 error = 0;
1975         else if (error == 0)
1976                 error = ENOTEMPTY;
1977         hammer_done_cursor(&cursor);
1978         return(error);
1979 }
1980