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 1.80 2008/07/01 02:08:58 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_add(hammer_record_t record);
  40 static int hammer_mem_lookup(hammer_cursor_t cursor);
  41 static int hammer_mem_first(hammer_cursor_t cursor);
  42 static int hammer_frontend_trunc_callback(hammer_record_t record,
  43                                 void *data __unused);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45
  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,
 256                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 257         record->flush_state = HAMMER_FST_IDLE;
 258         record->ip = ip;
 259         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 260         record->leaf.data_len = data_len;
 261         hammer_ref(&record->lock);
 262
 263         if (data_len) {
 264                 record->data = kmalloc(data_len, M_HAMMER, M_WAITOK | M_ZERO);
 265                 record->flags |= HAMMER_RECF_ALLOCDATA;
 266                 ++hammer_count_record_datas;
 267         }
 268
 269         return (record);
 270 }
 271
 272 void
 273 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 274 {
 275         while (record->flush_state == HAMMER_FST_FLUSH) {
 276                 record->flags |= HAMMER_RECF_WANTED;
 277                 tsleep(record, 0, ident, 0);
 278         }
 279 }
 280
 281 /*
 282  * Called from the backend, hammer_inode.c, after a record has been
 283  * flushed to disk.  The record has been exclusively locked by the
 284  * caller and interlocked with BE.
 285  *
 286  * We clean up the state, unlock, and release the record (the record
 287  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 288  */
 289 void
 290 hammer_flush_record_done(hammer_record_t record, int error)
 291 {
 292         hammer_inode_t target_ip;
 293
 294         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 295         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 296
 297         if (error) {
 298                 /*
 299                  * An error occured, the backend was unable to sync the
 300                  * record to its media.  Leave the record intact.
 301                  */
 302                 Debugger("flush_record_done error");
 303         }
 304
 305         if (record->flags & HAMMER_RECF_DELETED_BE) {
 306                 if ((target_ip = record->target_ip) != NULL) {
 307                         TAILQ_REMOVE(&target_ip->target_list, record,
 308                                      target_entry);
 309                         record->target_ip = NULL;
 310                         hammer_test_inode(target_ip);
 311                 }
 312                 record->flush_state = HAMMER_FST_IDLE;
 313         } else {
 314                 if (record->target_ip) {
 315                         record->flush_state = HAMMER_FST_SETUP;
 316                         hammer_test_inode(record->ip);
 317                         hammer_test_inode(record->target_ip);
 318                 } else {
 319                         record->flush_state = HAMMER_FST_IDLE;
 320                 }
 321         }
 322         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 323         if (record->flags & HAMMER_RECF_WANTED) {
 324                 record->flags &= ~HAMMER_RECF_WANTED;
 325                 wakeup(record);
 326         }
 327         hammer_rel_mem_record(record);
 328 }
 329
 330 /*
 331  * Release a memory record.  Records marked for deletion are immediately
 332  * removed from the RB-Tree but otherwise left intact until the last ref
 333  * goes away.
 334  */
 335 void
 336 hammer_rel_mem_record(struct hammer_record *record)
 337 {
 338         hammer_inode_t ip, target_ip;
 339
 340         hammer_unref(&record->lock);
 341
 342         if (record->lock.refs == 0) {
 343                 /*
 344                  * Upon release of the last reference wakeup any waiters.
 345                  * The record structure may get destroyed so callers will
 346                  * loop up and do a relookup.
 347                  *
 348                  * WARNING!  Record must be removed from RB-TREE before we
 349                  * might possibly block.  hammer_test_inode() can block!
 350                  */
 351                 ip = record->ip;
 352
 353                 /*
 354                  * Upon release of the last reference a record marked deleted
 355                  * is destroyed.
 356                  */
 357                 if (record->flags & HAMMER_RECF_DELETED_FE) {
 358                         KKASSERT(ip->lock.refs > 0);
 359                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 360
 361                         /*
 362                          * target_ip may have zero refs, we have to ref it
 363                          * to prevent it from being ripped out from under
 364                          * us.
 365                          */
 366                         if ((target_ip = record->target_ip) != NULL) {
 367                                 TAILQ_REMOVE(&target_ip->target_list,
 368                                              record, target_entry);
 369                                 record->target_ip = NULL;
 370                                 hammer_ref(&target_ip->lock);
 371                         }
 372
 373                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 374                                 RB_REMOVE(hammer_rec_rb_tree,
 375                                           &record->ip->rec_tree,
 376                                           record);
 377                                 KKASSERT(ip->rsv_recs > 0);
 378                                 --ip->hmp->rsv_recs;
 379                                 --ip->rsv_recs;
 380                                 ip->hmp->rsv_databytes -= record->leaf.data_len;
 381                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 382
 383                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 384                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 385                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 386                                         hammer_test_inode(record->ip);
 387                                 }
 388                         }
 389
 390                         /*
 391                          * Do this test after removing record from the B-Tree.
 392                          */
 393                         if (target_ip) {
 394                                 hammer_test_inode(target_ip);
 395                                 hammer_rel_inode(target_ip, 0);
 396                         }
 397
 398                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 399                                 --hammer_count_record_datas;
 400                                 kfree(record->data, M_HAMMER);
 401                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 402                         }
 403                         if (record->resv) {
 404                                 hammer_blockmap_reserve_complete(ip->hmp,
 405                                                                  record->resv);
 406                                 record->resv = NULL;
 407                         }
 408                         record->data = NULL;
 409                         --hammer_count_records;
 410                         kfree(record, M_HAMMER);
 411                 }
 412         }
 413 }
 414
 415 /*
 416  * Record visibility depends on whether the record is being accessed by
 417  * the backend or the frontend.
 418  *
 419  * Return non-zero if the record is visible, zero if it isn't or if it is
 420  * deleted.
 421  */
 422 static __inline
 423 int
 424 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 425 {
 426         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 427                 if (record->flags & HAMMER_RECF_DELETED_BE)
 428                         return(0);
 429         } else {
 430                 if (record->flags & HAMMER_RECF_DELETED_FE)
 431                         return(0);
 432         }
 433         return(1);
 434 }
 435
 436 /*
 437  * This callback is used as part of the RB_SCAN function for in-memory
 438  * records.  We terminate it (return -1) as soon as we get a match.
 439  *
 440  * This routine is used by frontend code.
 441  *
 442  * The primary compare code does not account for ASOF lookups.  This
 443  * code handles that case as well as a few others.
 444  */
 445 static
 446 int
 447 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 448 {
 449         hammer_cursor_t cursor = data;
 450
 451         /*
 452          * We terminate on success, so this should be NULL on entry.
 453          */
 454         KKASSERT(cursor->iprec == NULL);
 455
 456         /*
 457          * Skip if the record was marked deleted.
 458          */
 459         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 460                 return(0);
 461
 462         /*
 463          * Skip if not visible due to our as-of TID
 464          */
 465         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 466                 if (cursor->asof < rec->leaf.base.create_tid)
 467                         return(0);
 468                 if (rec->leaf.base.delete_tid &&
 469                     cursor->asof >= rec->leaf.base.delete_tid) {
 470                         return(0);
 471                 }
 472         }
 473
 474         /*
 475          * If the record is queued to the flusher we have to block until
 476          * it isn't.  Otherwise we may see duplication between our memory
 477          * cache and the media.
 478          */
 479         hammer_ref(&rec->lock);
 480
 481 #warning "This deadlocks"
 482 #if 0
 483         if (rec->flush_state == HAMMER_FST_FLUSH)
 484                 hammer_wait_mem_record(rec);
 485 #endif
 486
 487         /*
 488          * The record may have been deleted while we were blocked.
 489          */
 490         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 491                 hammer_rel_mem_record(rec);
 492                 return(0);
 493         }
 494
 495         /*
 496          * Set the matching record and stop the scan.
 497          */
 498         cursor->iprec = rec;
 499         return(-1);
 500 }
 501
 502
 503 /*
 504  * Lookup an in-memory record given the key specified in the cursor.  Works
 505  * just like hammer_btree_lookup() but operates on an inode's in-memory
 506  * record list.
 507  *
 508  * The lookup must fail if the record is marked for deferred deletion.
 509  */
 510 static
 511 int
 512 hammer_mem_lookup(hammer_cursor_t cursor)
 513 {
 514         int error;
 515
 516         KKASSERT(cursor->ip);
 517         if (cursor->iprec) {
 518                 hammer_rel_mem_record(cursor->iprec);
 519                 cursor->iprec = NULL;
 520         }
 521         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 522                                    hammer_rec_scan_callback, cursor);
 523
 524         if (cursor->iprec == NULL)
 525                 error = ENOENT;
 526         else
 527                 error = 0;
 528         return(error);
 529 }
 530
 531 /*
 532  * hammer_mem_first() - locate the first in-memory record matching the
 533  * cursor within the bounds of the key range.
 534  */
 535 static
 536 int
 537 hammer_mem_first(hammer_cursor_t cursor)
 538 {
 539         hammer_inode_t ip;
 540
 541         ip = cursor->ip;
 542         KKASSERT(ip != NULL);
 543
 544         if (cursor->iprec) {
 545                 hammer_rel_mem_record(cursor->iprec);
 546                 cursor->iprec = NULL;
 547         }
 548
 549         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 550                                    hammer_rec_scan_callback, cursor);
 551
 552         /*
 553          * Adjust scan.node and keep it linked into the RB-tree so we can
 554          * hold the cursor through third party modifications of the RB-tree.
 555          */
 556         if (cursor->iprec)
 557                 return(0);
 558         return(ENOENT);
 559 }
 560
 561 void
 562 hammer_mem_done(hammer_cursor_t cursor)
 563 {
 564         if (cursor->iprec) {
 565                 hammer_rel_mem_record(cursor->iprec);
 566                 cursor->iprec = NULL;
 567         }
 568 }
 569
 570 /************************************************************************
 571  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 572  ************************************************************************
 573  *
 574  * These functions manipulate in-memory records.  Such records typically
 575  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 576  */
 577
 578 /*
 579  * Add a directory entry (dip,ncp) which references inode (ip).
 580  *
 581  * Note that the low 32 bits of the namekey are set temporarily to create
 582  * a unique in-memory record, and may be modified a second time when the
 583  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 584  * all 0's when synching to disk, which is not handled here.
 585  *
 586  * NOTE: bytes does not include any terminating \0 on name, and name might
 587  * not be terminated.
 588  */
 589 int
 590 hammer_ip_add_directory(struct hammer_transaction *trans,
 591                      struct hammer_inode *dip, const char *name, int bytes,
 592                      struct hammer_inode *ip)
 593 {
 594         struct hammer_cursor cursor;
 595         hammer_record_t record;
 596         int error;
 597         int count;
 598         u_int32_t iterator;
 599
 600         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 601         if (++trans->hmp->namekey_iterator == 0)
 602                 ++trans->hmp->namekey_iterator;
 603
 604         record->type = HAMMER_MEM_RECORD_ADD;
 605         record->leaf.base.localization = dip->obj_localization +
 606                                          HAMMER_LOCALIZE_MISC;
 607         record->leaf.base.obj_id = dip->obj_id;
 608         record->leaf.base.key = hammer_directory_namekey(name, bytes);
 609         record->leaf.base.key += trans->hmp->namekey_iterator;
 610         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 611         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 612         record->data->entry.obj_id = ip->obj_id;
 613         record->data->entry.localization = ip->obj_localization;
 614         bcopy(name, record->data->entry.name, bytes);
 615
 616         ++ip->ino_data.nlinks;
 617         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 618
 619         /*
 620          * Find an unused namekey.  Both the in-memory record tree and
 621          * the B-Tree are checked.  Exact matches also match create_tid
 622          * so use an ASOF search to (mostly) ignore it.
 623          *
 624          * delete-visibility is set so pending deletions do not give us
 625          * a false-negative on our ability to use an iterator.
 626          */
 627         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 628         cursor.key_beg = record->leaf.base;
 629         cursor.flags |= HAMMER_CURSOR_ASOF;
 630         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 631         cursor.asof = ip->obj_asof;
 632
 633         count = 0;
 634         while (hammer_ip_lookup(&cursor) == 0) {
 635                 iterator = (u_int32_t)record->leaf.base.key + 1;
 636                 if (iterator == 0)
 637                         iterator = 1;
 638                 record->leaf.base.key &= ~0xFFFFFFFFLL;
 639                 record->leaf.base.key |= iterator;
 640                 cursor.key_beg.key = record->leaf.base.key;
 641                 if (++count == 1000000000) {
 642                         hammer_rel_mem_record(record);
 643                         error = ENOSPC;
 644                         goto failed;
 645                 }
 646         }
 647
 648         /*
 649          * The target inode and the directory entry are bound together.
 650          */
 651         record->target_ip = ip;
 652         record->flush_state = HAMMER_FST_SETUP;
 653         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 654
 655         /*
 656          * The inode now has a dependancy and must be taken out of the idle
 657          * state.  An inode not in an idle state is given an extra reference.
 658          */
 659         if (ip->flush_state == HAMMER_FST_IDLE) {
 660                 hammer_ref(&ip->lock);
 661                 ip->flush_state = HAMMER_FST_SETUP;
 662         }
 663         error = hammer_mem_add(record);
 664 failed:
 665         hammer_done_cursor(&cursor);
 666         return(error);
 667 }
 668
 669 /*
 670  * Delete the directory entry and update the inode link count.  The
 671  * cursor must be seeked to the directory entry record being deleted.
 672  *
 673  * The related inode should be share-locked by the caller.  The caller is
 674  * on the frontend.
 675  *
 676  * This function can return EDEADLK requiring the caller to terminate
 677  * the cursor, any locks, wait on the returned record, and retry.
 678  */
 679 int
 680 hammer_ip_del_directory(struct hammer_transaction *trans,
 681                      hammer_cursor_t cursor, struct hammer_inode *dip,
 682                      struct hammer_inode *ip)
 683 {
 684         hammer_record_t record;
 685         int error;
 686
 687         if (hammer_cursor_inmem(cursor)) {
 688                 /*
 689                  * In-memory (unsynchronized) records can simply be freed.
 690                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 691                  * by the backend, we must still avoid races against the
 692                  * backend potentially syncing the record to the media.
 693                  *
 694                  * We cannot call hammer_ip_delete_record(), that routine may
 695                  * only be called from the backend.
 696                  */
 697                 record = cursor->iprec;
 698                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 699                         KKASSERT(cursor->deadlk_rec == NULL);
 700                         hammer_ref(&record->lock);
 701                         cursor->deadlk_rec = record;
 702                         error = EDEADLK;
 703                 } else {
 704                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 705                         record->flags |= HAMMER_RECF_DELETED_FE;
 706                         error = 0;
 707                 }
 708         } else {
 709                 /*
 710                  * If the record is on-disk we have to queue the deletion by
 711                  * the record's key.  This also causes lookups to skip the
 712                  * record.
 713                  */
 714                 KKASSERT(dip->flags &
 715                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 716                 record = hammer_alloc_mem_record(dip, 0);
 717                 record->type = HAMMER_MEM_RECORD_DEL;
 718                 record->leaf.base = cursor->leaf->base;
 719
 720                 record->target_ip = ip;
 721                 record->flush_state = HAMMER_FST_SETUP;
 722                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 723
 724                 /*
 725                  * The inode now has a dependancy and must be taken out of
 726                  * the idle state.  An inode not in an idle state is given
 727                  * an extra reference.
 728                  */
 729                 if (ip->flush_state == HAMMER_FST_IDLE) {
 730                         hammer_ref(&ip->lock);
 731                         ip->flush_state = HAMMER_FST_SETUP;
 732                 }
 733
 734                 error = hammer_mem_add(record);
 735         }
 736
 737         /*
 738          * One less link.  The file may still be open in the OS even after
 739          * all links have gone away.
 740          *
 741          * We have to terminate the cursor before syncing the inode to
 742          * avoid deadlocking against ourselves.  XXX this may no longer
 743          * be true.
 744          *
 745          * If nlinks drops to zero and the vnode is inactive (or there is
 746          * no vnode), call hammer_inode_unloadable_check() to zonk the
 747          * inode.  If we don't do this here the inode will not be destroyed
 748          * on-media until we unmount.
 749          */
 750         if (error == 0) {
 751                 --ip->ino_data.nlinks;
 752                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 753                 if (ip->ino_data.nlinks == 0 &&
 754                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 755                         hammer_done_cursor(cursor);
 756                         hammer_inode_unloadable_check(ip, 1);
 757                         hammer_flush_inode(ip, 0);
 758                 }
 759
 760         }
 761         return(error);
 762 }
 763
 764 /*
 765  * Add a record to an inode.
 766  *
 767  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 768  * initialize the following additional fields:
 769  *
 770  * The related inode should be share-locked by the caller.  The caller is
 771  * on the frontend.
 772  *
 773  * record->rec.entry.base.base.key
 774  * record->rec.entry.base.base.rec_type
 775  * record->rec.entry.base.base.data_len
 776  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 777  */
 778 int
 779 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 780 {
 781         hammer_inode_t ip = record->ip;
 782         int error;
 783
 784         KKASSERT(record->leaf.base.localization != 0);
 785         record->leaf.base.obj_id = ip->obj_id;
 786         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 787         error = hammer_mem_add(record);
 788         return(error);
 789 }
 790
 791 /*
 792  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 793  * reserved so the front-end can flush large data writes without having
 794  * to queue the BIO to the flusher.  Only the related record gets queued
 795  * to the flusher.
 796  */
 797 static hammer_record_t
 798 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 799 {
 800         hammer_record_t record;
 801         struct hammer_btree_leaf_elm leaf;
 802
 803         bzero(&leaf, sizeof(leaf));
 804         leaf.base.obj_id = ip->obj_id;
 805         leaf.base.key = file_offset + bytes;
 806         leaf.base.create_tid = 0;
 807         leaf.base.delete_tid = 0;
 808         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 809         leaf.base.obj_type = 0;                 /* unused */
 810         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
 811         leaf.base.localization = ip->obj_localization + HAMMER_LOCALIZE_MISC;
 812         leaf.data_len = bytes;
 813
 814         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
 815         if (record)
 816                 hammer_ref(&record->lock);
 817         return(record);
 818 }
 819
 820 /*
 821  * Reserve blockmap space placemarked with an in-memory record.
 822  *
 823  * This routine is called by the frontend in order to be able to directly
 824  * flush a buffer cache buffer.  The frontend has locked the related buffer
 825  * cache buffers and we should be able to manipulate any overlapping
 826  * in-memory records.
 827  */
 828 hammer_record_t
 829 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 830                    int *errorp)
 831 {
 832         hammer_record_t record;
 833         hammer_record_t conflict;
 834         int zone;
 835         int flags;
 836
 837         /*
 838          * Deal with conflicting in-memory records.  We cannot have multiple
 839          * in-memory records for the same offset without seriously confusing
 840          * the backend, including but not limited to the backend issuing
 841          * delete-create-delete sequences and asserting on the delete_tid
 842          * being the same as the create_tid.
 843          *
 844          * If we encounter a record with the backend interlock set we cannot
 845          * immediately delete it without confusing the backend.
 846          */
 847         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 848                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 849                         conflict->flags |= HAMMER_RECF_WANTED;
 850                         tsleep(conflict, 0, "hmrrc3", 0);
 851                 } else {
 852                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 853                 }
 854                 hammer_rel_mem_record(conflict);
 855         }
 856
 857         /*
 858          * Create a record to cover the direct write.  This is called with
 859          * the related BIO locked so there should be no possible conflict.
 860          *
 861          * The backend is responsible for finalizing the space reserved in
 862          * this record.
 863          *
 864          * XXX bytes not aligned, depend on the reservation code to
 865          * align the reservation.
 866          */
 867         record = hammer_alloc_mem_record(ip, 0);
 868         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 869                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 870         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 871                                                &record->leaf.data_offset,
 872                                                errorp);
 873         if (record->resv == NULL) {
 874                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 875                 hammer_rel_mem_record(record);
 876                 return(NULL);
 877         }
 878         record->type = HAMMER_MEM_RECORD_DATA;
 879         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 880         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 881         record->leaf.base.obj_id = ip->obj_id;
 882         record->leaf.base.key = file_offset + bytes;
 883         record->leaf.base.localization = ip->obj_localization +
 884                                          HAMMER_LOCALIZE_MISC;
 885         record->leaf.data_len = bytes;
 886         hammer_crc_set_leaf(data, &record->leaf);
 887         flags = record->flags;
 888
 889         hammer_ref(&record->lock);      /* mem_add eats a reference */
 890         *errorp = hammer_mem_add(record);
 891         if (*errorp) {
 892                 conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 893                 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
 894                         *errorp, conflict, file_offset, bytes);
 895                 if (conflict)
 896                         kprintf("conflict %lld %d\n", conflict->leaf.base.key, conflict->leaf.data_len);
 897                 if (conflict)
 898                         hammer_rel_mem_record(conflict);
 899         }
 900         KKASSERT(*errorp == 0);
 901         conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 902         if (conflict != record) {
 903                 kprintf("conflict mismatch %p %p %08x\n", conflict, record, record->flags);
 904                 if (conflict)
 905                     kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict->leaf.base.key, conflict->leaf.data_len, record->leaf.base.key, record->leaf.data_len);
 906         }
 907         KKASSERT(conflict == record);
 908         hammer_rel_mem_record(conflict);
 909
 910         return (record);
 911 }
 912
 913 /*
 914  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 915  * and records in a flushing state are handled by the backend.  The vnops
 916  * setattr code will handle the block containing the truncation point.
 917  *
 918  * Partial blocks are not deleted.
 919  */
 920 int
 921 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 922 {
 923         struct rec_trunc_info info;
 924
 925         switch(ip->ino_data.obj_type) {
 926         case HAMMER_OBJTYPE_REGFILE:
 927                 info.rec_type = HAMMER_RECTYPE_DATA;
 928                 break;
 929         case HAMMER_OBJTYPE_DBFILE:
 930                 info.rec_type = HAMMER_RECTYPE_DB;
 931                 break;
 932         default:
 933                 return(EINVAL);
 934         }
 935         info.trunc_off = file_size;
 936         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 937                                    hammer_frontend_trunc_callback, &info);
 938         return(0);
 939 }
 940
 941 static int
 942 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
 943 {
 944         if (record->flags & HAMMER_RECF_DELETED_FE)
 945                 return(0);
 946         if (record->flush_state == HAMMER_FST_FLUSH)
 947                 return(0);
 948         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 949         hammer_ref(&record->lock);
 950         record->flags |= HAMMER_RECF_DELETED_FE;
 951         hammer_rel_mem_record(record);
 952         return(0);
 953 }
 954
 955 /*
 956  * Return 1 if the caller must check for and delete existing records
 957  * before writing out a new data record.
 958  *
 959  * Return 0 if the caller can just insert the record into the B-Tree without
 960  * checking.
 961  */
 962 static int
 963 hammer_record_needs_overwrite_delete(hammer_record_t record)
 964 {
 965         hammer_inode_t ip = record->ip;
 966         int64_t file_offset;
 967         int r;
 968
 969         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
 970                 file_offset = record->leaf.base.key;
 971         else
 972                 file_offset = record->leaf.base.key - record->leaf.data_len;
 973         r = (file_offset < ip->save_trunc_off);
 974         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
 975                 if (ip->save_trunc_off <= record->leaf.base.key)
 976                         ip->save_trunc_off = record->leaf.base.key + 1;
 977         } else {
 978                 if (ip->save_trunc_off < record->leaf.base.key)
 979                         ip->save_trunc_off = record->leaf.base.key;
 980         }
 981         return(r);
 982 }
 983
 984 /*
 985  * Backend code.  Sync a record to the media.
 986  */
 987 int
 988 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
 989 {
 990         hammer_transaction_t trans = cursor->trans;
 991         int64_t file_offset;
 992         int bytes;
 993         void *bdata;
 994         int error;
 995
 996         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 997         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 998         KKASSERT(record->leaf.base.localization != 0);
 999
1000         /*
1001          * If this is a bulk-data record placemarker there may be an existing
1002          * record on-disk, indicating a data overwrite.  If there is the
1003          * on-disk record must be deleted before we can insert our new record.
1004          *
1005          * We've synthesized this record and do not know what the create_tid
1006          * on-disk is, nor how much data it represents.
1007          *
1008          * Keep in mind that (key) for data records is (base_offset + len),
1009          * not (base_offset).  Also, we only want to get rid of on-disk
1010          * records since we are trying to sync our in-memory record, call
1011          * hammer_ip_delete_range() with truncating set to 1 to make sure
1012          * it skips in-memory records.
1013          *
1014          * It is ok for the lookup to return ENOENT.
1015          *
1016          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1017          * to call hammer_ip_delete_range() or not.  This also means we must
1018          * update sync_trunc_off() as we write.
1019          */
1020         if (record->type == HAMMER_MEM_RECORD_DATA &&
1021             hammer_record_needs_overwrite_delete(record)) {
1022                 file_offset = record->leaf.base.key - record->leaf.data_len;
1023                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1024                         ~HAMMER_BUFMASK;
1025                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1026                 error = hammer_ip_delete_range(
1027                                 cursor, record->ip,
1028                                 file_offset, file_offset + bytes - 1,
1029                                 1);
1030                 if (error && error != ENOENT)
1031                         goto done;
1032         }
1033
1034         /*
1035          * Setup the cursor.
1036          */
1037         hammer_normalize_cursor(cursor);
1038         cursor->key_beg = record->leaf.base;
1039         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1040         cursor->flags |= HAMMER_CURSOR_BACKEND;
1041         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1042
1043         /*
1044          * Records can wind up on-media before the inode itself is on-media.
1045          * Flag the case.
1046          */
1047         record->ip->flags |= HAMMER_INODE_DONDISK;
1048
1049         /*
1050          * If we are deleting a directory entry an exact match must be
1051          * found on-disk.
1052          */
1053         if (record->type == HAMMER_MEM_RECORD_DEL) {
1054                 error = hammer_btree_lookup(cursor);
1055                 if (error == 0) {
1056                         error = hammer_ip_delete_record(cursor, record->ip,
1057                                                         trans->tid);
1058                         if (error == 0) {
1059                                 record->flags |= HAMMER_RECF_DELETED_FE;
1060                                 record->flags |= HAMMER_RECF_DELETED_BE;
1061                         }
1062                 }
1063                 goto done;
1064         }
1065
1066         /*
1067          * We are inserting.
1068          *
1069          * Issue a lookup to position the cursor and locate the cluster.  The
1070          * target key should not exist.  If we are creating a directory entry
1071          * we may have to iterate the low 32 bits of the key to find an unused
1072          * key.
1073          */
1074         cursor->flags |= HAMMER_CURSOR_INSERT;
1075
1076         error = hammer_btree_lookup(cursor);
1077         if (hammer_debug_inode)
1078                 kprintf("DOINSERT LOOKUP %d\n", error);
1079         if (error == 0) {
1080                 kprintf("hammer_ip_sync_record: duplicate rec "
1081                         "at (%016llx)\n", record->leaf.base.key);
1082                 Debugger("duplicate record1");
1083                 error = EIO;
1084         }
1085 #if 0
1086         if (record->type == HAMMER_MEM_RECORD_DATA)
1087                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1088                         record->leaf.base.key - record->leaf.data_len,
1089                         record->leaf.data_offset, error);
1090 #endif
1091
1092         if (error != ENOENT)
1093                 goto done;
1094
1095         /*
1096          * Allocate the record and data.  The result buffers will be
1097          * marked as being modified and further calls to
1098          * hammer_modify_buffer() will result in unneeded UNDO records.
1099          *
1100          * Support zero-fill records (data == NULL and data_len != 0)
1101          */
1102         if (record->type == HAMMER_MEM_RECORD_DATA) {
1103                 /*
1104                  * The data portion of a bulk-data record has already been
1105                  * committed to disk, we need only adjust the layer2
1106                  * statistics in the same transaction as our B-Tree insert.
1107                  */
1108                 KKASSERT(record->leaf.data_offset != 0);
1109                 hammer_blockmap_finalize(trans, record->leaf.data_offset,
1110                                          record->leaf.data_len);
1111                 error = 0;
1112         } else if (record->data && record->leaf.data_len) {
1113                 /*
1114                  * Wholely cached record, with data.  Allocate the data.
1115                  */
1116                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1117                                           record->leaf.base.rec_type,
1118                                           &record->leaf.data_offset,
1119                                           &cursor->data_buffer, &error);
1120                 if (bdata == NULL)
1121                         goto done;
1122                 hammer_crc_set_leaf(record->data, &record->leaf);
1123                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1124                 bcopy(record->data, bdata, record->leaf.data_len);
1125                 hammer_modify_buffer_done(cursor->data_buffer);
1126         } else {
1127                 /*
1128                  * Wholely cached record, without data.
1129                  */
1130                 record->leaf.data_offset = 0;
1131                 record->leaf.data_crc = 0;
1132         }
1133
1134         error = hammer_btree_insert(cursor, &record->leaf);
1135         if (hammer_debug_inode && error)
1136                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1137
1138         /*
1139          * Our record is on-disk, normally mark the in-memory version as
1140          * deleted.  If the record represented a directory deletion but
1141          * we had to sync a valid directory entry to disk we must convert
1142          * the record to a covering delete so the frontend does not have
1143          * visibility on the synced entry.
1144          */
1145         if (error == 0) {
1146                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1147                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1148                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1149                         record->type = HAMMER_MEM_RECORD_DEL;
1150                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1151                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1152                         /* hammer_flush_record_done takes care of the rest */
1153                 } else {
1154                         record->flags |= HAMMER_RECF_DELETED_FE;
1155                         record->flags |= HAMMER_RECF_DELETED_BE;
1156                 }
1157         } else {
1158                 if (record->leaf.data_offset) {
1159                         hammer_blockmap_free(trans, record->leaf.data_offset,
1160                                              record->leaf.data_len);
1161                 }
1162         }
1163
1164 done:
1165         return(error);
1166 }
1167
1168 /*
1169  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1170  * entry's key is used to deal with hash collisions in the upper 32 bits.
1171  * A unique 64 bit key is generated in-memory and may be regenerated a
1172  * second time when the directory record is flushed to the on-disk B-Tree.
1173  *
1174  * A referenced record is passed to this function.  This function
1175  * eats the reference.  If an error occurs the record will be deleted.
1176  *
1177  * A copy of the temporary record->data pointer provided by the caller
1178  * will be made.
1179  */
1180 static
1181 int
1182 hammer_mem_add(hammer_record_t record)
1183 {
1184         hammer_mount_t hmp = record->ip->hmp;
1185
1186         /*
1187          * Make a private copy of record->data
1188          */
1189         if (record->data)
1190                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1191
1192         /*
1193          * Insert into the RB tree.  A unique key should have already
1194          * been selected if this is a directory entry.
1195          */
1196         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1197                 record->flags |= HAMMER_RECF_DELETED_FE;
1198                 hammer_rel_mem_record(record);
1199                 return (EEXIST);
1200         }
1201         ++hmp->count_newrecords;
1202         ++hmp->rsv_recs;
1203         ++record->ip->rsv_recs;
1204         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1205         record->flags |= HAMMER_RECF_ONRBTREE;
1206         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1207         hammer_rel_mem_record(record);
1208         return(0);
1209 }
1210
1211 /************************************************************************
1212  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1213  ************************************************************************
1214  *
1215  * These functions augment the B-Tree scanning functions in hammer_btree.c
1216  * by merging in-memory records with on-disk records.
1217  */
1218
1219 /*
1220  * Locate a particular record either in-memory or on-disk.
1221  *
1222  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1223  * NOT be called to iterate results.
1224  */
1225 int
1226 hammer_ip_lookup(hammer_cursor_t cursor)
1227 {
1228         int error;
1229
1230         /*
1231          * If the element is in-memory return it without searching the
1232          * on-disk B-Tree
1233          */
1234         KKASSERT(cursor->ip);
1235         error = hammer_mem_lookup(cursor);
1236         if (error == 0) {
1237                 cursor->leaf = &cursor->iprec->leaf;
1238                 return(error);
1239         }
1240         if (error != ENOENT)
1241                 return(error);
1242
1243         /*
1244          * If the inode has on-disk components search the on-disk B-Tree.
1245          */
1246         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1247                 return(error);
1248         error = hammer_btree_lookup(cursor);
1249         if (error == 0)
1250                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1251         return(error);
1252 }
1253
1254 /*
1255  * Locate the first record within the cursor's key_beg/key_end range,
1256  * restricted to a particular inode.  0 is returned on success, ENOENT
1257  * if no records matched the requested range, or some other error.
1258  *
1259  * When 0 is returned hammer_ip_next() may be used to iterate additional
1260  * records within the requested range.
1261  *
1262  * This function can return EDEADLK, requiring the caller to terminate
1263  * the cursor and try again.
1264  */
1265 int
1266 hammer_ip_first(hammer_cursor_t cursor)
1267 {
1268         hammer_inode_t ip = cursor->ip;
1269         int error;
1270
1271         KKASSERT(ip != NULL);
1272
1273         /*
1274          * Clean up fields and setup for merged scan
1275          */
1276         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1277         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1278         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1279         if (cursor->iprec) {
1280                 hammer_rel_mem_record(cursor->iprec);
1281                 cursor->iprec = NULL;
1282         }
1283
1284         /*
1285          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1286          * exact lookup so if we get ENOENT we have to call the iterate
1287          * function to validate the first record after the begin key.
1288          *
1289          * The ATEDISK flag is used by hammer_btree_iterate to determine
1290          * whether it must index forwards or not.  It is also used here
1291          * to select the next record from in-memory or on-disk.
1292          *
1293          * EDEADLK can only occur if the lookup hit an empty internal
1294          * element and couldn't delete it.  Since this could only occur
1295          * in-range, we can just iterate from the failure point.
1296          */
1297         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1298                 error = hammer_btree_lookup(cursor);
1299                 if (error == ENOENT || error == EDEADLK) {
1300                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1301                         if (hammer_debug_general & 0x2000)
1302                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1303                         error = hammer_btree_iterate(cursor);
1304                 }
1305                 if (error && error != ENOENT)
1306                         return(error);
1307                 if (error == 0) {
1308                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1309                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1310                 } else {
1311                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1312                 }
1313         }
1314
1315         /*
1316          * Search the in-memory record list (Red-Black tree).  Unlike the
1317          * B-Tree search, mem_first checks for records in the range.
1318          */
1319         error = hammer_mem_first(cursor);
1320         if (error && error != ENOENT)
1321                 return(error);
1322         if (error == 0) {
1323                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1324                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1325                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1326                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1327         }
1328
1329         /*
1330          * This will return the first matching record.
1331          */
1332         return(hammer_ip_next(cursor));
1333 }
1334
1335 /*
1336  * Retrieve the next record in a merged iteration within the bounds of the
1337  * cursor.  This call may be made multiple times after the cursor has been
1338  * initially searched with hammer_ip_first().
1339  *
1340  * 0 is returned on success, ENOENT if no further records match the
1341  * requested range, or some other error code is returned.
1342  */
1343 int
1344 hammer_ip_next(hammer_cursor_t cursor)
1345 {
1346         hammer_btree_elm_t elm;
1347         hammer_record_t rec, save;
1348         int error;
1349         int r;
1350
1351 next_btree:
1352         /*
1353          * Load the current on-disk and in-memory record.  If we ate any
1354          * records we have to get the next one.
1355          *
1356          * If we deleted the last on-disk record we had scanned ATEDISK will
1357          * be clear and DELBTREE will be set, forcing a call to iterate. The
1358          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1359          * element.  If ATEDISK is set, iterate will skip the 'current'
1360          * element.
1361          *
1362          * Get the next on-disk record
1363          */
1364         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1365                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1366                         error = hammer_btree_iterate(cursor);
1367                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1368                         if (error == 0) {
1369                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1370                                 hammer_cache_node(&cursor->ip->cache[1],
1371                                                   cursor->node);
1372                         } else {
1373                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1374                                                  HAMMER_CURSOR_ATEDISK;
1375                         }
1376                 }
1377         }
1378
1379 next_memory:
1380         /*
1381          * Get the next in-memory record.  The record can be ripped out
1382          * of the RB tree so we maintain a scan_info structure to track
1383          * the next node.
1384          *
1385          * hammer_rec_scan_cmp:  Is the record still in our general range,
1386          *                       (non-inclusive of snapshot exclusions)?
1387          * hammer_rec_scan_callback: Is the record in our snapshot?
1388          */
1389         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1390                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1391                         save = cursor->iprec;
1392                         cursor->iprec = NULL;
1393                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1394                         while (rec) {
1395                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1396                                         break;
1397                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1398                                         break;
1399                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1400                         }
1401                         if (save)
1402                                 hammer_rel_mem_record(save);
1403                         if (cursor->iprec) {
1404                                 KKASSERT(cursor->iprec == rec);
1405                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1406                         } else {
1407                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1408                         }
1409                 }
1410         }
1411
1412         /*
1413          * The memory record may have become stale while being held in
1414          * cursor->iprec.  We are interlocked against the backend on
1415          * with regards to B-Tree entries.
1416          */
1417         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1418                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1419                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1420                         goto next_memory;
1421                 }
1422         }
1423
1424         /*
1425          * Extract either the disk or memory record depending on their
1426          * relative position.
1427          */
1428         error = 0;
1429         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1430         case 0:
1431                 /*
1432                  * Both entries valid.   Compare the entries and nominally
1433                  * return the first one in the sort order.  Numerous cases
1434                  * require special attention, however.
1435                  */
1436                 elm = &cursor->node->ondisk->elms[cursor->index];
1437                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1438
1439                 /*
1440                  * If the two entries differ only by their key (-2/2) or
1441                  * create_tid (-1/1), and are DATA records, we may have a
1442                  * nominal match.  We have to calculate the base file
1443                  * offset of the data.
1444                  */
1445                 if (r <= 2 && r >= -2 && r != 0 &&
1446                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1447                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1448                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1449                         int64_t base2 = cursor->iprec->leaf.base.key -
1450                                         cursor->iprec->leaf.data_len;
1451                         if (base1 == base2)
1452                                 r = 0;
1453                 }
1454
1455                 if (r < 0) {
1456                         error = hammer_btree_extract(cursor,
1457                                                      HAMMER_CURSOR_GET_LEAF);
1458                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1459                         break;
1460                 }
1461
1462                 /*
1463                  * If the entries match exactly the memory entry is either
1464                  * an on-disk directory entry deletion or a bulk data
1465                  * overwrite.  If it is a directory entry deletion we eat
1466                  * both entries.
1467                  *
1468                  * For the bulk-data overwrite case it is possible to have
1469                  * visibility into both, which simply means the syncer
1470                  * hasn't gotten around to doing the delete+insert sequence
1471                  * on the B-Tree.  Use the memory entry and throw away the
1472                  * on-disk entry.
1473                  *
1474                  * If the in-memory record is not either of these we
1475                  * probably caught the syncer while it was syncing it to
1476                  * the media.  Since we hold a shared lock on the cursor,
1477                  * the in-memory record had better be marked deleted at
1478                  * this point.
1479                  */
1480                 if (r == 0) {
1481                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1482                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1483                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1484                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1485                                         goto next_btree;
1486                                 }
1487                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1488                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1489                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1490                                 }
1491                                 /* fall through to memory entry */
1492                         } else {
1493                                 panic("hammer_ip_next: duplicate mem/b-tree entry");
1494                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1495                                 goto next_memory;
1496                         }
1497                 }
1498                 /* fall through to the memory entry */
1499         case HAMMER_CURSOR_ATEDISK:
1500                 /*
1501                  * Only the memory entry is valid.
1502                  */
1503                 cursor->leaf = &cursor->iprec->leaf;
1504                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1505
1506                 /*
1507                  * If the memory entry is an on-disk deletion we should have
1508                  * also had found a B-Tree record.  If the backend beat us
1509                  * to it it would have interlocked the cursor and we should
1510                  * have seen the in-memory record marked DELETED_FE.
1511                  */
1512                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1513                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1514                         panic("hammer_ip_next: del-on-disk with no b-tree entry");
1515                 }
1516                 break;
1517         case HAMMER_CURSOR_ATEMEM:
1518                 /*
1519                  * Only the disk entry is valid
1520                  */
1521                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1522                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1523                 break;
1524         default:
1525                 /*
1526                  * Neither entry is valid
1527                  *
1528                  * XXX error not set properly
1529                  */
1530                 cursor->leaf = NULL;
1531                 error = ENOENT;
1532                 break;
1533         }
1534         return(error);
1535 }
1536
1537 /*
1538  * Resolve the cursor->data pointer for the current cursor position in
1539  * a merged iteration.
1540  */
1541 int
1542 hammer_ip_resolve_data(hammer_cursor_t cursor)
1543 {
1544         hammer_record_t record;
1545         int error;
1546
1547         if (hammer_cursor_inmem(cursor)) {
1548                 /*
1549                  * The data associated with an in-memory record is usually
1550                  * kmalloced, but reserve-ahead data records will have an
1551                  * on-disk reference.
1552                  *
1553                  * NOTE: Reserve-ahead data records must be handled in the
1554                  * context of the related high level buffer cache buffer
1555                  * to interlock against async writes.
1556                  */
1557                 record = cursor->iprec;
1558                 cursor->data = record->data;
1559                 error = 0;
1560                 if (cursor->data == NULL) {
1561                         KKASSERT(record->leaf.base.rec_type ==
1562                                  HAMMER_RECTYPE_DATA);
1563                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1564                                                     record->leaf.data_offset,
1565                                                     record->leaf.data_len,
1566                                                     &error,
1567                                                     &cursor->data_buffer);
1568                 }
1569         } else {
1570                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1571                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1572         }
1573         return(error);
1574 }
1575
1576 /*
1577  * Backend truncation / record replacement - delete records in range.
1578  *
1579  * Delete all records within the specified range for inode ip.  In-memory
1580  * records still associated with the frontend are ignored.
1581  *
1582  * If truncating is non-zero in-memory records associated with the back-end
1583  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1584  *
1585  * NOTES:
1586  *
1587  *      * An unaligned range will cause new records to be added to cover
1588  *        the edge cases. (XXX not implemented yet).
1589  *
1590  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1591  *        also do not deal with unaligned ranges.
1592  *
1593  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1594  *
1595  *      * Record keys for regular file data have to be special-cased since
1596  *        they indicate the end of the range (key = base + bytes).
1597  *
1598  *      * This function may be asked to delete ridiculously huge ranges, for
1599  *        example if someone truncates or removes a 1TB regular file.  We
1600  *        must be very careful on restarts and we may have to stop w/
1601  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1602  */
1603 int
1604 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1605                        int64_t ran_beg, int64_t ran_end, int truncating)
1606 {
1607         hammer_transaction_t trans = cursor->trans;
1608         hammer_btree_leaf_elm_t leaf;
1609         int error;
1610         int64_t off;
1611         int64_t tmp64;
1612
1613 #if 0
1614         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1615 #endif
1616
1617         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1618 retry:
1619         hammer_normalize_cursor(cursor);
1620         cursor->key_beg.localization = ip->obj_localization +
1621                                        HAMMER_LOCALIZE_MISC;
1622         cursor->key_beg.obj_id = ip->obj_id;
1623         cursor->key_beg.create_tid = 0;
1624         cursor->key_beg.delete_tid = 0;
1625         cursor->key_beg.obj_type = 0;
1626
1627         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1628                 cursor->key_beg.key = ran_beg;
1629                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1630         } else {
1631                 /*
1632                  * The key in the B-Tree is (base+bytes), so the first possible
1633                  * matching key is ran_beg + 1.
1634                  */
1635                 cursor->key_beg.key = ran_beg + 1;
1636                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1637         }
1638
1639         cursor->key_end = cursor->key_beg;
1640         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1641                 cursor->key_end.key = ran_end;
1642         } else {
1643                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1644                 if (tmp64 < ran_end)
1645                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1646                 else
1647                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1648         }
1649
1650         cursor->asof = ip->obj_asof;
1651         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1652         cursor->flags |= HAMMER_CURSOR_ASOF;
1653         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1654         cursor->flags |= HAMMER_CURSOR_BACKEND;
1655         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1656
1657         error = hammer_ip_first(cursor);
1658
1659         /*
1660          * Iterate through matching records and mark them as deleted.
1661          */
1662         while (error == 0) {
1663                 leaf = cursor->leaf;
1664
1665                 KKASSERT(leaf->base.delete_tid == 0);
1666
1667                 /*
1668                  * There may be overlap cases for regular file data.  Also
1669                  * remember the key for a regular file record is (base + len),
1670                  * NOT (base).
1671                  */
1672                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1673                         off = leaf->base.key - leaf->data_len;
1674                         /*
1675                          * Check the left edge case.  We currently do not
1676                          * split existing records.
1677                          */
1678                         if (off < ran_beg) {
1679                                 panic("hammer left edge case %016llx %d\n",
1680                                         leaf->base.key, leaf->data_len);
1681                         }
1682
1683                         /*
1684                          * Check the right edge case.  Note that the
1685                          * record can be completely out of bounds, which
1686                          * terminates the search.
1687                          *
1688                          * base->key is exclusive of the right edge while
1689                          * ran_end is inclusive of the right edge.  The
1690                          * (key - data_len) left boundary is inclusive.
1691                          *
1692                          * XXX theory-check this test at some point, are
1693                          * we missing a + 1 somewhere?  Note that ran_end
1694                          * could overflow.
1695                          */
1696                         if (leaf->base.key - 1 > ran_end) {
1697                                 if (leaf->base.key - leaf->data_len > ran_end)
1698                                         break;
1699                                 panic("hammer right edge case\n");
1700                         }
1701                 } else {
1702                         off = leaf->base.key;
1703                 }
1704
1705                 /*
1706                  * Delete the record.  When truncating we do not delete
1707                  * in-memory (data) records because they represent data
1708                  * written after the truncation.
1709                  *
1710                  * This will also physically destroy the B-Tree entry and
1711                  * data if the retention policy dictates.  The function
1712                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1713                  * uses to perform a fixup.
1714                  */
1715                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1716                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1717                         /*
1718                          * If we have built up too many meta-buffers we risk
1719                          * deadlocking the kernel and must stop.  This can
1720                          * occur when deleting ridiculously huge files.
1721                          * sync_trunc_off is updated so the next cycle does
1722                          * not re-iterate records we have already deleted.
1723                          *
1724                          * This is only done with formal truncations.
1725                          */
1726                         if (truncating > 1 && error == 0 &&
1727                             hammer_flusher_meta_limit(ip->hmp)) {
1728                                 ip->sync_trunc_off = off;
1729                                 error = EWOULDBLOCK;
1730                         }
1731                 }
1732                 if (error)
1733                         break;
1734                 ran_beg = off;  /* for restart */
1735                 error = hammer_ip_next(cursor);
1736         }
1737         if (cursor->node)
1738                 hammer_cache_node(&ip->cache[1], cursor->node);
1739
1740         if (error == EDEADLK) {
1741                 hammer_done_cursor(cursor);
1742                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1743                 if (error == 0)
1744                         goto retry;
1745         }
1746         if (error == ENOENT)
1747                 error = 0;
1748         return(error);
1749 }
1750
1751 /*
1752  * This function deletes remaining auxillary records when an inode is
1753  * being deleted.  This function explicitly does not delete the
1754  * inode record, directory entry, data, or db records.  Those must be
1755  * properly disposed of prior to this call.
1756  */
1757 int
1758 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
1759 {
1760         hammer_transaction_t trans = cursor->trans;
1761         hammer_btree_leaf_elm_t leaf;
1762         int error;
1763
1764         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1765 retry:
1766         hammer_normalize_cursor(cursor);
1767         cursor->key_beg.localization = ip->obj_localization +
1768                                        HAMMER_LOCALIZE_MISC;
1769         cursor->key_beg.obj_id = ip->obj_id;
1770         cursor->key_beg.create_tid = 0;
1771         cursor->key_beg.delete_tid = 0;
1772         cursor->key_beg.obj_type = 0;
1773         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
1774         cursor->key_beg.key = HAMMER_MIN_KEY;
1775
1776         cursor->key_end = cursor->key_beg;
1777         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
1778         cursor->key_end.key = HAMMER_MAX_KEY;
1779
1780         cursor->asof = ip->obj_asof;
1781         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1782         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1783         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1784         cursor->flags |= HAMMER_CURSOR_BACKEND;
1785
1786         error = hammer_ip_first(cursor);
1787
1788         /*
1789          * Iterate through matching records and mark them as deleted.
1790          */
1791         while (error == 0) {
1792                 leaf = cursor->leaf;
1793
1794                 KKASSERT(leaf->base.delete_tid == 0);
1795
1796                 /*
1797                  * Mark the record and B-Tree entry as deleted.  This will
1798                  * also physically delete the B-Tree entry, record, and
1799                  * data if the retention policy dictates.  The function
1800                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1801                  * uses to perform a fixup.
1802                  *
1803                  * Directory entries (and delete-on-disk directory entries)
1804                  * must be synced and cannot be deleted.
1805                  */
1806                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1807                 ++*countp;
1808                 if (error)
1809                         break;
1810                 error = hammer_ip_next(cursor);
1811         }
1812         if (cursor->node)
1813                 hammer_cache_node(&ip->cache[1], cursor->node);
1814         if (error == EDEADLK) {
1815                 hammer_done_cursor(cursor);
1816                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1817                 if (error == 0)
1818                         goto retry;
1819         }
1820         if (error == ENOENT)
1821                 error = 0;
1822         return(error);
1823 }
1824
1825 /*
1826  * Delete the record at the current cursor.  On success the cursor will
1827  * be positioned appropriately for an iteration but may no longer be at
1828  * a leaf node.
1829  *
1830  * This routine is only called from the backend.
1831  *
1832  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1833  * cursor and retry.
1834  */
1835 int
1836 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1837                         hammer_tid_t tid)
1838 {
1839         hammer_off_t zone2_offset;
1840         hammer_record_t iprec;
1841         hammer_btree_elm_t elm;
1842         hammer_mount_t hmp;
1843         int error;
1844         int dodelete;
1845
1846         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1847         KKASSERT(tid != 0);
1848         hmp = cursor->node->hmp;
1849
1850         /*
1851          * In-memory (unsynchronized) records can simply be freed.  This
1852          * only occurs in range iterations since all other records are
1853          * individually synchronized.  Thus there should be no confusion with
1854          * the interlock.
1855          *
1856          * An in-memory record may be deleted before being committed to disk,
1857          * but could have been accessed in the mean time.  The backing store
1858          * may never been marked allocated and so hammer_blockmap_free() may
1859          * never get called on it.  Because of this we have to make sure that
1860          * we've gotten rid of any related hammer_buffer or buffer cache
1861          * buffer.
1862          */
1863         if (hammer_cursor_inmem(cursor)) {
1864                 iprec = cursor->iprec;
1865                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1866                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1867                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1868
1869                 if (iprec->leaf.data_offset && iprec->leaf.data_len) {
1870                         zone2_offset = hammer_blockmap_lookup(hmp, iprec->leaf.data_offset, &error);
1871                         KKASSERT(error == 0);
1872                         hammer_del_buffers(hmp,
1873                                            iprec->leaf.data_offset,
1874                                            zone2_offset,
1875                                            iprec->leaf.data_len);
1876                 }
1877                 return(0);
1878         }
1879
1880         /*
1881          * On-disk records are marked as deleted by updating their delete_tid.
1882          * This does not effect their position in the B-Tree (which is based
1883          * on their create_tid).
1884          */
1885         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1886         elm = NULL;
1887
1888         /*
1889          * If we were mounted with the nohistory option, we physically
1890          * delete the record.
1891          */
1892         dodelete = hammer_nohistory(ip);
1893
1894         if (error == 0) {
1895                 error = hammer_cursor_upgrade(cursor);
1896                 if (error == 0) {
1897                         elm = &cursor->node->ondisk->elms[cursor->index];
1898                         hammer_modify_node(cursor->trans, cursor->node,
1899                                            elm, sizeof(*elm));
1900                         elm->leaf.base.delete_tid = tid;
1901                         elm->leaf.delete_ts = cursor->trans->time32;
1902                         hammer_modify_node_done(cursor->node);
1903
1904                         /*
1905                          * An on-disk record cannot have the same delete_tid
1906                          * as its create_tid.  In a chain of record updates
1907                          * this could result in a duplicate record.
1908                          */
1909                         KKASSERT(elm->leaf.base.delete_tid != elm->leaf.base.create_tid);
1910                 }
1911         }
1912
1913         if (error == 0 && dodelete) {
1914                 error = hammer_delete_at_cursor(cursor, NULL);
1915                 if (error) {
1916                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1917                         error = 0;
1918                 }
1919         }
1920         return(error);
1921 }
1922
1923 int
1924 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
1925 {
1926         hammer_btree_elm_t elm;
1927         hammer_off_t data_offset;
1928         int32_t data_len;
1929         u_int16_t rec_type;
1930         int error;
1931
1932         elm = &cursor->node->ondisk->elms[cursor->index];
1933         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1934
1935         data_offset = elm->leaf.data_offset;
1936         data_len = elm->leaf.data_len;
1937         rec_type = elm->leaf.base.rec_type;
1938
1939         error = hammer_btree_delete(cursor);
1940         if (error == 0) {
1941                 /*
1942                  * This forces a fixup for the iteration because
1943                  * the cursor is now either sitting at the 'next'
1944                  * element or sitting at the end of a leaf.
1945                  */
1946                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1947                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
1948                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1949                 }
1950         }
1951         if (error == 0) {
1952                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
1953                 case HAMMER_ZONE_LARGE_DATA:
1954                 case HAMMER_ZONE_SMALL_DATA:
1955                 case HAMMER_ZONE_META:
1956                         hammer_blockmap_free(cursor->trans,
1957                                              data_offset, data_len);
1958                         break;
1959                 default:
1960                         break;
1961                 }
1962         }
1963         return (error);
1964 }
1965
1966 /*
1967  * Determine whether we can remove a directory.  This routine checks whether
1968  * a directory is empty or not and enforces flush connectivity.
1969  *
1970  * Flush connectivity requires that we block if the target directory is
1971  * currently flushing, otherwise it may not end up in the same flush group.
1972  *
1973  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
1974  */
1975 int
1976 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
1977 {
1978         struct hammer_cursor cursor;
1979         int error;
1980
1981         /*
1982          * Check directory empty
1983          */
1984         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
1985
1986         cursor.key_beg.localization = ip->obj_localization +
1987                                       HAMMER_LOCALIZE_MISC;
1988         cursor.key_beg.obj_id = ip->obj_id;
1989         cursor.key_beg.create_tid = 0;
1990         cursor.key_beg.delete_tid = 0;
1991         cursor.key_beg.obj_type = 0;
1992         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1993         cursor.key_beg.key = HAMMER_MIN_KEY;
1994
1995         cursor.key_end = cursor.key_beg;
1996         cursor.key_end.rec_type = 0xFFFF;
1997         cursor.key_end.key = HAMMER_MAX_KEY;
1998
1999         cursor.asof = ip->obj_asof;
2000         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2001
2002         error = hammer_ip_first(&cursor);
2003         if (error == ENOENT)
2004                 error = 0;
2005         else if (error == 0)
2006                 error = ENOTEMPTY;
2007         hammer_done_cursor(&cursor);
2008         return(error);
2009 }
2010