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 1.97 2008/09/23 22:28:56 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_lookup(hammer_cursor_t cursor);
  40 static int hammer_mem_first(hammer_cursor_t cursor);
  41 static int hammer_frontend_trunc_callback(hammer_record_t record,
  42                                 void *data __unused);
  43 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
  46                       hammer_btree_leaf_elm_t leaf);
  47
  48 struct rec_trunc_info {
  49         u_int16_t       rec_type;
  50         int64_t         trunc_off;
  51 };
  52
  53 struct hammer_bulk_info {
  54         hammer_record_t record;
  55         struct hammer_btree_leaf_elm leaf;
  56 };
  57
  58 /*
  59  * Red-black tree support.  Comparison code for insertion.
  60  */
  61 static int
  62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  63 {
  64         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  65                 return(-1);
  66         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  67                 return(1);
  68
  69         if (rec1->leaf.base.key < rec2->leaf.base.key)
  70                 return(-1);
  71         if (rec1->leaf.base.key > rec2->leaf.base.key)
  72                 return(1);
  73
  74         /*
  75          * Never match against an item deleted by the front-end.
  76          *
  77          * rec1 is greater then rec2 if rec1 is marked deleted.
  78          * rec1 is less then rec2 if rec2 is marked deleted.
  79          *
  80          * Multiple deleted records may be present, do not return 0
  81          * if both are marked deleted.
  82          */
  83         if (rec1->flags & HAMMER_RECF_DELETED_FE)
  84                 return(1);
  85         if (rec2->flags & HAMMER_RECF_DELETED_FE)
  86                 return(-1);
  87
  88         return(0);
  89 }
  90
  91 /*
  92  * Basic record comparison code similar to hammer_btree_cmp().
  93  */
  94 static int
  95 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
  96 {
  97         if (elm->rec_type < rec->leaf.base.rec_type)
  98                 return(-3);
  99         if (elm->rec_type > rec->leaf.base.rec_type)
 100                 return(3);
 101
 102         if (elm->key < rec->leaf.base.key)
 103                 return(-2);
 104         if (elm->key > rec->leaf.base.key)
 105                 return(2);
 106
 107         /*
 108          * Never match against an item deleted by the front-end.
 109          * elm is less then rec if rec is marked deleted.
 110          */
 111         if (rec->flags & HAMMER_RECF_DELETED_FE)
 112                 return(-1);
 113         return(0);
 114 }
 115
 116 /*
 117  * Ranged scan to locate overlapping record(s).  This is used by
 118  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 119  * to use a ranged scan because the keys for data records with the
 120  * same file base offset can be different due to differing data_len's.
 121  *
 122  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 123  */
 124 static int
 125 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 126 {
 127         struct hammer_bulk_info *info = data;
 128         hammer_btree_leaf_elm_t leaf = &info->leaf;
 129
 130         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 131                 return(-3);
 132         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 133                 return(3);
 134
 135         /*
 136          * Overlap compare
 137          */
 138         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 139                 /* rec_beg >= leaf_end */
 140                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 141                         return(2);
 142                 /* rec_end <= leaf_beg */
 143                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 144                         return(-2);
 145         } else {
 146                 if (rec->leaf.base.key < leaf->base.key)
 147                         return(-2);
 148                 if (rec->leaf.base.key > leaf->base.key)
 149                         return(2);
 150         }
 151
 152         /*
 153          * We have to return 0 at this point, even if DELETED_FE is set,
 154          * because returning anything else will cause the scan to ignore
 155          * one of the branches when we really want it to check both.
 156          */
 157         return(0);
 158 }
 159
 160 /*
 161  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 162  * is reversed so the comparison result has to be negated.  key_beg and
 163  * key_end are both range-inclusive.
 164  *
 165  * Localized deletions are not cached in-memory.
 166  */
 167 static
 168 int
 169 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 170 {
 171         hammer_cursor_t cursor = data;
 172         int r;
 173
 174         r = hammer_rec_cmp(&cursor->key_beg, rec);
 175         if (r > 1)
 176                 return(-1);
 177         r = hammer_rec_cmp(&cursor->key_end, rec);
 178         if (r < -1)
 179                 return(1);
 180         return(0);
 181 }
 182
 183 /*
 184  * This compare function is used when simply looking up key_beg.
 185  */
 186 static
 187 int
 188 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 189 {
 190         hammer_cursor_t cursor = data;
 191         int r;
 192
 193         r = hammer_rec_cmp(&cursor->key_beg, rec);
 194         if (r > 1)
 195                 return(-1);
 196         if (r < -1)
 197                 return(1);
 198         return(0);
 199 }
 200
 201 /*
 202  * Locate blocks within the truncation range.  Partial blocks do not count.
 203  */
 204 static
 205 int
 206 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 207 {
 208         struct rec_trunc_info *info = data;
 209
 210         if (rec->leaf.base.rec_type < info->rec_type)
 211                 return(-1);
 212         if (rec->leaf.base.rec_type > info->rec_type)
 213                 return(1);
 214
 215         switch(rec->leaf.base.rec_type) {
 216         case HAMMER_RECTYPE_DB:
 217                 /*
 218                  * DB record key is not beyond the truncation point, retain.
 219                  */
 220                 if (rec->leaf.base.key < info->trunc_off)
 221                         return(-1);
 222                 break;
 223         case HAMMER_RECTYPE_DATA:
 224                 /*
 225                  * DATA record offset start is not beyond the truncation point,
 226                  * retain.
 227                  */
 228                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 229                         return(-1);
 230                 break;
 231         default:
 232                 panic("hammer_rec_trunc_cmp: unexpected record type");
 233         }
 234
 235         /*
 236          * The record start is >= the truncation point, return match,
 237          * the record should be destroyed.
 238          */
 239         return(0);
 240 }
 241
 242 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 243
 244 /*
 245  * Allocate a record for the caller to finish filling in.  The record is
 246  * returned referenced.
 247  */
 248 hammer_record_t
 249 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 250 {
 251         hammer_record_t record;
 252         hammer_mount_t hmp;
 253
 254         hmp = ip->hmp;
 255         ++hammer_count_records;
 256         record = kmalloc(sizeof(*record), hmp->m_misc,
 257                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 258         record->flush_state = HAMMER_FST_IDLE;
 259         record->ip = ip;
 260         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 261         record->leaf.data_len = data_len;
 262         hammer_ref(&record->lock);
 263
 264         if (data_len) {
 265                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
 266                 record->flags |= HAMMER_RECF_ALLOCDATA;
 267                 ++hammer_count_record_datas;
 268         }
 269
 270         return (record);
 271 }
 272
 273 void
 274 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 275 {
 276         while (record->flush_state == HAMMER_FST_FLUSH) {
 277                 record->flags |= HAMMER_RECF_WANTED;
 278                 tsleep(record, 0, ident, 0);
 279         }
 280 }
 281
 282 /*
 283  * Called from the backend, hammer_inode.c, after a record has been
 284  * flushed to disk.  The record has been exclusively locked by the
 285  * caller and interlocked with BE.
 286  *
 287  * We clean up the state, unlock, and release the record (the record
 288  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 289  */
 290 void
 291 hammer_flush_record_done(hammer_record_t record, int error)
 292 {
 293         hammer_inode_t target_ip;
 294
 295         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 296         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 297
 298         if (error) {
 299                 /*
 300                  * An error occured, the backend was unable to sync the
 301                  * record to its media.  Leave the record intact.
 302                  */
 303                 hammer_critical_error(record->ip->hmp, record->ip, error,
 304                                       "while flushing record");
 305         }
 306
 307         --record->flush_group->refs;
 308         record->flush_group = NULL;
 309
 310         if (record->flags & HAMMER_RECF_DELETED_BE) {
 311                 if ((target_ip = record->target_ip) != NULL) {
 312                         TAILQ_REMOVE(&target_ip->target_list, record,
 313                                      target_entry);
 314                         record->target_ip = NULL;
 315                         hammer_test_inode(target_ip);
 316                 }
 317                 record->flush_state = HAMMER_FST_IDLE;
 318         } else {
 319                 if (record->target_ip) {
 320                         record->flush_state = HAMMER_FST_SETUP;
 321                         hammer_test_inode(record->ip);
 322                         hammer_test_inode(record->target_ip);
 323                 } else {
 324                         record->flush_state = HAMMER_FST_IDLE;
 325                 }
 326         }
 327         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 328         if (record->flags & HAMMER_RECF_WANTED) {
 329                 record->flags &= ~HAMMER_RECF_WANTED;
 330                 wakeup(record);
 331         }
 332         hammer_rel_mem_record(record);
 333 }
 334
 335 /*
 336  * Release a memory record.  Records marked for deletion are immediately
 337  * removed from the RB-Tree but otherwise left intact until the last ref
 338  * goes away.
 339  */
 340 void
 341 hammer_rel_mem_record(struct hammer_record *record)
 342 {
 343         hammer_mount_t hmp;
 344         hammer_reserve_t resv;
 345         hammer_inode_t ip;
 346         hammer_inode_t target_ip;
 347
 348         hammer_unref(&record->lock);
 349
 350         if (record->lock.refs == 0) {
 351                 /*
 352                  * Upon release of the last reference wakeup any waiters.
 353                  * The record structure may get destroyed so callers will
 354                  * loop up and do a relookup.
 355                  *
 356                  * WARNING!  Record must be removed from RB-TREE before we
 357                  * might possibly block.  hammer_test_inode() can block!
 358                  */
 359                 ip = record->ip;
 360                 hmp = ip->hmp;
 361
 362                 /*
 363                  * Upon release of the last reference a record marked deleted
 364                  * is destroyed.
 365                  */
 366                 if (record->flags & HAMMER_RECF_DELETED_FE) {
 367                         KKASSERT(ip->lock.refs > 0);
 368                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 369
 370                         /*
 371                          * target_ip may have zero refs, we have to ref it
 372                          * to prevent it from being ripped out from under
 373                          * us.
 374                          */
 375                         if ((target_ip = record->target_ip) != NULL) {
 376                                 TAILQ_REMOVE(&target_ip->target_list,
 377                                              record, target_entry);
 378                                 record->target_ip = NULL;
 379                                 hammer_ref(&target_ip->lock);
 380                         }
 381
 382                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 383                                 RB_REMOVE(hammer_rec_rb_tree,
 384                                           &record->ip->rec_tree,
 385                                           record);
 386                                 KKASSERT(ip->rsv_recs > 0);
 387                                 --hmp->rsv_recs;
 388                                 --ip->rsv_recs;
 389                                 hmp->rsv_databytes -= record->leaf.data_len;
 390                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 391
 392                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 393                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 394                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 395                                         hammer_test_inode(record->ip);
 396                                 }
 397                         }
 398
 399                         /*
 400                          * We must wait for any direct-IO to complete before
 401                          * we can destroy the record because the bio may
 402                          * have a reference to it.
 403                          */
 404                         if (record->flags &
 405                            (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL)) {
 406                                 hammer_io_direct_wait(record);
 407                         }
 408
 409
 410                         /*
 411                          * Do this test after removing record from the B-Tree.
 412                          */
 413                         if (target_ip) {
 414                                 hammer_test_inode(target_ip);
 415                                 hammer_rel_inode(target_ip, 0);
 416                         }
 417
 418                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 419                                 --hammer_count_record_datas;
 420                                 kfree(record->data, hmp->m_misc);
 421                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 422                         }
 423
 424                         /*
 425                          * Release the reservation.  If the record was not
 426                          * committed return the reservation before
 427                          * releasing it.
 428                          */
 429                         if ((resv = record->resv) != NULL) {
 430 #if 0
 431                                 if ((record->flags & HAMMER_RECF_COMMITTED) == 0) {
 432                                         hammer_blockmap_reserve_undo(
 433                                                 hmp, resv,
 434                                                 record->leaf.data_offset,
 435                                                 record->leaf.data_len);
 436                                 }
 437 #endif
 438                                 hammer_blockmap_reserve_complete(hmp, resv);
 439                                 record->resv = NULL;
 440                         }
 441                         record->data = NULL;
 442                         --hammer_count_records;
 443                         kfree(record, hmp->m_misc);
 444                 }
 445         }
 446 }
 447
 448 /*
 449  * Record visibility depends on whether the record is being accessed by
 450  * the backend or the frontend.
 451  *
 452  * Return non-zero if the record is visible, zero if it isn't or if it is
 453  * deleted.
 454  *
 455  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 456  * records to be returned.  This is so pending deletions are detected
 457  * when using an iterator to locate an unused hash key, or when we need
 458  * to locate historical records on-disk to destroy.
 459  */
 460 static __inline
 461 int
 462 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 463 {
 464         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 465                 return(1);
 466         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 467                 if (record->flags & HAMMER_RECF_DELETED_BE)
 468                         return(0);
 469         } else {
 470                 if (record->flags & HAMMER_RECF_DELETED_FE)
 471                         return(0);
 472         }
 473         return(1);
 474 }
 475
 476 /*
 477  * This callback is used as part of the RB_SCAN function for in-memory
 478  * records.  We terminate it (return -1) as soon as we get a match.
 479  *
 480  * This routine is used by frontend code.
 481  *
 482  * The primary compare code does not account for ASOF lookups.  This
 483  * code handles that case as well as a few others.
 484  */
 485 static
 486 int
 487 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 488 {
 489         hammer_cursor_t cursor = data;
 490
 491         /*
 492          * We terminate on success, so this should be NULL on entry.
 493          */
 494         KKASSERT(cursor->iprec == NULL);
 495
 496         /*
 497          * Skip if the record was marked deleted.
 498          */
 499         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 500                 return(0);
 501
 502         /*
 503          * Skip if not visible due to our as-of TID
 504          */
 505         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 506                 if (cursor->asof < rec->leaf.base.create_tid)
 507                         return(0);
 508                 if (rec->leaf.base.delete_tid &&
 509                     cursor->asof >= rec->leaf.base.delete_tid) {
 510                         return(0);
 511                 }
 512         }
 513
 514         /*
 515          * ref the record.  The record is protected from backend B-Tree
 516          * interactions by virtue of the cursor's IP lock.
 517          */
 518         hammer_ref(&rec->lock);
 519
 520         /*
 521          * The record may have been deleted while we were blocked.
 522          */
 523         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 524                 hammer_rel_mem_record(rec);
 525                 return(0);
 526         }
 527
 528         /*
 529          * Set the matching record and stop the scan.
 530          */
 531         cursor->iprec = rec;
 532         return(-1);
 533 }
 534
 535
 536 /*
 537  * Lookup an in-memory record given the key specified in the cursor.  Works
 538  * just like hammer_btree_lookup() but operates on an inode's in-memory
 539  * record list.
 540  *
 541  * The lookup must fail if the record is marked for deferred deletion.
 542  */
 543 static
 544 int
 545 hammer_mem_lookup(hammer_cursor_t cursor)
 546 {
 547         int error;
 548
 549         KKASSERT(cursor->ip);
 550         if (cursor->iprec) {
 551                 hammer_rel_mem_record(cursor->iprec);
 552                 cursor->iprec = NULL;
 553         }
 554         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 555                                    hammer_rec_scan_callback, cursor);
 556
 557         if (cursor->iprec == NULL)
 558                 error = ENOENT;
 559         else
 560                 error = 0;
 561         return(error);
 562 }
 563
 564 /*
 565  * hammer_mem_first() - locate the first in-memory record matching the
 566  * cursor within the bounds of the key range.
 567  */
 568 static
 569 int
 570 hammer_mem_first(hammer_cursor_t cursor)
 571 {
 572         hammer_inode_t ip;
 573
 574         ip = cursor->ip;
 575         KKASSERT(ip != NULL);
 576
 577         if (cursor->iprec) {
 578                 hammer_rel_mem_record(cursor->iprec);
 579                 cursor->iprec = NULL;
 580         }
 581
 582         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 583                                    hammer_rec_scan_callback, cursor);
 584
 585         /*
 586          * Adjust scan.node and keep it linked into the RB-tree so we can
 587          * hold the cursor through third party modifications of the RB-tree.
 588          */
 589         if (cursor->iprec)
 590                 return(0);
 591         return(ENOENT);
 592 }
 593
 594 /************************************************************************
 595  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 596  ************************************************************************
 597  *
 598  * These functions manipulate in-memory records.  Such records typically
 599  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 600  */
 601
 602 /*
 603  * Add a directory entry (dip,ncp) which references inode (ip).
 604  *
 605  * Note that the low 32 bits of the namekey are set temporarily to create
 606  * a unique in-memory record, and may be modified a second time when the
 607  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 608  * all 0's when synching to disk, which is not handled here.
 609  *
 610  * NOTE: bytes does not include any terminating \0 on name, and name might
 611  * not be terminated.
 612  */
 613 int
 614 hammer_ip_add_directory(struct hammer_transaction *trans,
 615                      struct hammer_inode *dip, const char *name, int bytes,
 616                      struct hammer_inode *ip)
 617 {
 618         struct hammer_cursor cursor;
 619         hammer_record_t record;
 620         int error;
 621         u_int32_t max_iterations;
 622
 623         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 624
 625         record->type = HAMMER_MEM_RECORD_ADD;
 626         record->leaf.base.localization = dip->obj_localization +
 627                                          HAMMER_LOCALIZE_MISC;
 628         record->leaf.base.obj_id = dip->obj_id;
 629         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 630                                                          &max_iterations);
 631         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 632         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 633         record->data->entry.obj_id = ip->obj_id;
 634         record->data->entry.localization = ip->obj_localization;
 635         bcopy(name, record->data->entry.name, bytes);
 636
 637         ++ip->ino_data.nlinks;
 638         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 639
 640         /*
 641          * Find an unused namekey.  Both the in-memory record tree and
 642          * the B-Tree are checked.  We do not want historically deleted
 643          * names to create a collision as our iteration space may be limited,
 644          * and since create_tid wouldn't match anyway an ASOF search
 645          * must be used to locate collisions.
 646          *
 647          * delete-visibility is set so pending deletions do not give us
 648          * a false-negative on our ability to use an iterator.
 649          *
 650          * The iterator must not rollover the key.  Directory keys only
 651          * use the positive key space.
 652          */
 653         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 654         cursor.key_beg = record->leaf.base;
 655         cursor.flags |= HAMMER_CURSOR_ASOF;
 656         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 657         cursor.asof = ip->obj_asof;
 658
 659         while (hammer_ip_lookup(&cursor) == 0) {
 660                 ++record->leaf.base.key;
 661                 KKASSERT(record->leaf.base.key > 0);
 662                 cursor.key_beg.key = record->leaf.base.key;
 663                 if (--max_iterations == 0) {
 664                         hammer_rel_mem_record(record);
 665                         error = ENOSPC;
 666                         goto failed;
 667                 }
 668         }
 669
 670         /*
 671          * The target inode and the directory entry are bound together.
 672          */
 673         record->target_ip = ip;
 674         record->flush_state = HAMMER_FST_SETUP;
 675         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 676
 677         /*
 678          * The inode now has a dependancy and must be taken out of the idle
 679          * state.  An inode not in an idle state is given an extra reference.
 680          *
 681          * When transitioning to a SETUP state flag for an automatic reflush
 682          * when the dependancies are disposed of if someone is waiting on
 683          * the inode.
 684          */
 685         if (ip->flush_state == HAMMER_FST_IDLE) {
 686                 hammer_ref(&ip->lock);
 687                 ip->flush_state = HAMMER_FST_SETUP;
 688                 if (ip->flags & HAMMER_INODE_FLUSHW)
 689                         ip->flags |= HAMMER_INODE_REFLUSH;
 690         }
 691         error = hammer_mem_add(record);
 692         if (error == 0) {
 693                 dip->ino_data.mtime = trans->time;
 694                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 695         }
 696 failed:
 697         hammer_done_cursor(&cursor);
 698         return(error);
 699 }
 700
 701 /*
 702  * Delete the directory entry and update the inode link count.  The
 703  * cursor must be seeked to the directory entry record being deleted.
 704  *
 705  * The related inode should be share-locked by the caller.  The caller is
 706  * on the frontend.
 707  *
 708  * This function can return EDEADLK requiring the caller to terminate
 709  * the cursor, any locks, wait on the returned record, and retry.
 710  */
 711 int
 712 hammer_ip_del_directory(struct hammer_transaction *trans,
 713                      hammer_cursor_t cursor, struct hammer_inode *dip,
 714                      struct hammer_inode *ip)
 715 {
 716         hammer_record_t record;
 717         int error;
 718
 719         if (hammer_cursor_inmem(cursor)) {
 720                 /*
 721                  * In-memory (unsynchronized) records can simply be freed.
 722                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 723                  * by the backend, we must still avoid races against the
 724                  * backend potentially syncing the record to the media.
 725                  *
 726                  * We cannot call hammer_ip_delete_record(), that routine may
 727                  * only be called from the backend.
 728                  */
 729                 record = cursor->iprec;
 730                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 731                         KKASSERT(cursor->deadlk_rec == NULL);
 732                         hammer_ref(&record->lock);
 733                         cursor->deadlk_rec = record;
 734                         error = EDEADLK;
 735                 } else {
 736                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 737                         record->flags |= HAMMER_RECF_DELETED_FE;
 738                         error = 0;
 739                 }
 740         } else {
 741                 /*
 742                  * If the record is on-disk we have to queue the deletion by
 743                  * the record's key.  This also causes lookups to skip the
 744                  * record.
 745                  */
 746                 KKASSERT(dip->flags &
 747                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 748                 record = hammer_alloc_mem_record(dip, 0);
 749                 record->type = HAMMER_MEM_RECORD_DEL;
 750                 record->leaf.base = cursor->leaf->base;
 751
 752                 record->target_ip = ip;
 753                 record->flush_state = HAMMER_FST_SETUP;
 754                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 755
 756                 /*
 757                  * The inode now has a dependancy and must be taken out of
 758                  * the idle state.  An inode not in an idle state is given
 759                  * an extra reference.
 760                  *
 761                  * When transitioning to a SETUP state flag for an automatic
 762                  * reflush when the dependancies are disposed of if someone
 763                  * is waiting on the inode.
 764                  */
 765                 if (ip->flush_state == HAMMER_FST_IDLE) {
 766                         hammer_ref(&ip->lock);
 767                         ip->flush_state = HAMMER_FST_SETUP;
 768                         if (ip->flags & HAMMER_INODE_FLUSHW)
 769                                 ip->flags |= HAMMER_INODE_REFLUSH;
 770                 }
 771
 772                 error = hammer_mem_add(record);
 773         }
 774
 775         /*
 776          * One less link.  The file may still be open in the OS even after
 777          * all links have gone away.
 778          *
 779          * We have to terminate the cursor before syncing the inode to
 780          * avoid deadlocking against ourselves.  XXX this may no longer
 781          * be true.
 782          *
 783          * If nlinks drops to zero and the vnode is inactive (or there is
 784          * no vnode), call hammer_inode_unloadable_check() to zonk the
 785          * inode.  If we don't do this here the inode will not be destroyed
 786          * on-media until we unmount.
 787          */
 788         if (error == 0) {
 789                 --ip->ino_data.nlinks;
 790                 dip->ino_data.mtime = trans->time;
 791                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 792                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 793                 if (ip->ino_data.nlinks == 0 &&
 794                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 795                         hammer_done_cursor(cursor);
 796                         hammer_inode_unloadable_check(ip, 1);
 797                         hammer_flush_inode(ip, 0);
 798                 }
 799
 800         }
 801         return(error);
 802 }
 803
 804 /*
 805  * Add a record to an inode.
 806  *
 807  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 808  * initialize the following additional fields:
 809  *
 810  * The related inode should be share-locked by the caller.  The caller is
 811  * on the frontend.
 812  *
 813  * record->rec.entry.base.base.key
 814  * record->rec.entry.base.base.rec_type
 815  * record->rec.entry.base.base.data_len
 816  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 817  */
 818 int
 819 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 820 {
 821         hammer_inode_t ip = record->ip;
 822         int error;
 823
 824         KKASSERT(record->leaf.base.localization != 0);
 825         record->leaf.base.obj_id = ip->obj_id;
 826         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 827         error = hammer_mem_add(record);
 828         return(error);
 829 }
 830
 831 /*
 832  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 833  * reserved so the front-end can flush large data writes without having
 834  * to queue the BIO to the flusher.  Only the related record gets queued
 835  * to the flusher.
 836  */
 837
 838 static hammer_record_t
 839 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 840 {
 841         struct hammer_bulk_info info;
 842
 843         bzero(&info, sizeof(info));
 844         info.leaf.base.obj_id = ip->obj_id;
 845         info.leaf.base.key = file_offset + bytes;
 846         info.leaf.base.create_tid = 0;
 847         info.leaf.base.delete_tid = 0;
 848         info.leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 849         info.leaf.base.obj_type = 0;                            /* unused */
 850         info.leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;        /* unused */
 851         info.leaf.base.localization = ip->obj_localization +    /* unused */
 852                                       HAMMER_LOCALIZE_MISC;
 853         info.leaf.data_len = bytes;
 854
 855         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 856                                    hammer_bulk_scan_callback, &info);
 857
 858         return(info.record);    /* may be NULL */
 859 }
 860
 861 /*
 862  * Take records vetted by overlap_cmp.  The first non-deleted record
 863  * (if any) stops the scan.
 864  */
 865 static int
 866 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 867 {
 868         struct hammer_bulk_info *info = data;
 869
 870         if (record->flags & HAMMER_RECF_DELETED_FE)
 871                 return(0);
 872         hammer_ref(&record->lock);
 873         info->record = record;
 874         return(-1);                     /* stop scan */
 875 }
 876
 877 /*
 878  * Reserve blockmap space placemarked with an in-memory record.
 879  *
 880  * This routine is called by the frontend in order to be able to directly
 881  * flush a buffer cache buffer.  The frontend has locked the related buffer
 882  * cache buffers and we should be able to manipulate any overlapping
 883  * in-memory records.
 884  *
 885  * The caller is responsible for adding the returned record.
 886  */
 887 hammer_record_t
 888 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 889                    int *errorp)
 890 {
 891         hammer_record_t record;
 892         hammer_record_t conflict;
 893         int zone;
 894
 895         /*
 896          * Deal with conflicting in-memory records.  We cannot have multiple
 897          * in-memory records for the same base offset without seriously
 898          * confusing the backend, including but not limited to the backend
 899          * issuing delete-create-delete or create-delete-create sequences
 900          * and asserting on the delete_tid being the same as the create_tid.
 901          *
 902          * If we encounter a record with the backend interlock set we cannot
 903          * immediately delete it without confusing the backend.
 904          */
 905         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 906                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 907                         conflict->flags |= HAMMER_RECF_WANTED;
 908                         tsleep(conflict, 0, "hmrrc3", 0);
 909                 } else {
 910                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 911                 }
 912                 hammer_rel_mem_record(conflict);
 913         }
 914
 915         /*
 916          * Create a record to cover the direct write.  This is called with
 917          * the related BIO locked so there should be no possible conflict.
 918          *
 919          * The backend is responsible for finalizing the space reserved in
 920          * this record.
 921          *
 922          * XXX bytes not aligned, depend on the reservation code to
 923          * align the reservation.
 924          */
 925         record = hammer_alloc_mem_record(ip, 0);
 926         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 927                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 928         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 929                                                &record->leaf.data_offset,
 930                                                errorp);
 931         if (record->resv == NULL) {
 932                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 933                 hammer_rel_mem_record(record);
 934                 return(NULL);
 935         }
 936         record->type = HAMMER_MEM_RECORD_DATA;
 937         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 938         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 939         record->leaf.base.obj_id = ip->obj_id;
 940         record->leaf.base.key = file_offset + bytes;
 941         record->leaf.base.localization = ip->obj_localization +
 942                                          HAMMER_LOCALIZE_MISC;
 943         record->leaf.data_len = bytes;
 944         hammer_crc_set_leaf(data, &record->leaf);
 945         KKASSERT(*errorp == 0);
 946         return(record);
 947 }
 948
 949 /*
 950  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 951  * and records in a flushing state are handled by the backend.  The vnops
 952  * setattr code will handle the block containing the truncation point.
 953  *
 954  * Partial blocks are not deleted.
 955  */
 956 int
 957 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 958 {
 959         struct rec_trunc_info info;
 960
 961         switch(ip->ino_data.obj_type) {
 962         case HAMMER_OBJTYPE_REGFILE:
 963                 info.rec_type = HAMMER_RECTYPE_DATA;
 964                 break;
 965         case HAMMER_OBJTYPE_DBFILE:
 966                 info.rec_type = HAMMER_RECTYPE_DB;
 967                 break;
 968         default:
 969                 return(EINVAL);
 970         }
 971         info.trunc_off = file_size;
 972         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 973                                    hammer_frontend_trunc_callback, &info);
 974         return(0);
 975 }
 976
 977 static int
 978 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
 979 {
 980         if (record->flags & HAMMER_RECF_DELETED_FE)
 981                 return(0);
 982         if (record->flush_state == HAMMER_FST_FLUSH)
 983                 return(0);
 984         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 985         hammer_ref(&record->lock);
 986         record->flags |= HAMMER_RECF_DELETED_FE;
 987         hammer_rel_mem_record(record);
 988         return(0);
 989 }
 990
 991 /*
 992  * Return 1 if the caller must check for and delete existing records
 993  * before writing out a new data record.
 994  *
 995  * Return 0 if the caller can just insert the record into the B-Tree without
 996  * checking.
 997  */
 998 static int
 999 hammer_record_needs_overwrite_delete(hammer_record_t record)
1000 {
1001         hammer_inode_t ip = record->ip;
1002         int64_t file_offset;
1003         int r;
1004
1005         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1006                 file_offset = record->leaf.base.key;
1007         else
1008                 file_offset = record->leaf.base.key - record->leaf.data_len;
1009         r = (file_offset < ip->save_trunc_off);
1010         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1011                 if (ip->save_trunc_off <= record->leaf.base.key)
1012                         ip->save_trunc_off = record->leaf.base.key + 1;
1013         } else {
1014                 if (ip->save_trunc_off < record->leaf.base.key)
1015                         ip->save_trunc_off = record->leaf.base.key;
1016         }
1017         return(r);
1018 }
1019
1020 /*
1021  * Backend code.  Sync a record to the media.
1022  */
1023 int
1024 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1025 {
1026         hammer_transaction_t trans = cursor->trans;
1027         int64_t file_offset;
1028         int bytes;
1029         void *bdata;
1030         int error;
1031         int doprop;
1032
1033         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1034         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1035         KKASSERT(record->leaf.base.localization != 0);
1036
1037         /*
1038          * Any direct-write related to the record must complete before we
1039          * can sync the record to the on-disk media.
1040          */
1041         if (record->flags & (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL))
1042                 hammer_io_direct_wait(record);
1043
1044         /*
1045          * If this is a bulk-data record placemarker there may be an existing
1046          * record on-disk, indicating a data overwrite.  If there is the
1047          * on-disk record must be deleted before we can insert our new record.
1048          *
1049          * We've synthesized this record and do not know what the create_tid
1050          * on-disk is, nor how much data it represents.
1051          *
1052          * Keep in mind that (key) for data records is (base_offset + len),
1053          * not (base_offset).  Also, we only want to get rid of on-disk
1054          * records since we are trying to sync our in-memory record, call
1055          * hammer_ip_delete_range() with truncating set to 1 to make sure
1056          * it skips in-memory records.
1057          *
1058          * It is ok for the lookup to return ENOENT.
1059          *
1060          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1061          * to call hammer_ip_delete_range() or not.  This also means we must
1062          * update sync_trunc_off() as we write.
1063          */
1064         if (record->type == HAMMER_MEM_RECORD_DATA &&
1065             hammer_record_needs_overwrite_delete(record)) {
1066                 file_offset = record->leaf.base.key - record->leaf.data_len;
1067                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1068                         ~HAMMER_BUFMASK;
1069                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1070                 error = hammer_ip_delete_range(
1071                                 cursor, record->ip,
1072                                 file_offset, file_offset + bytes - 1,
1073                                 1);
1074                 if (error && error != ENOENT)
1075                         goto done;
1076         }
1077
1078         /*
1079          * If this is a general record there may be an on-disk version
1080          * that must be deleted before we can insert the new record.
1081          */
1082         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1083                 error = hammer_delete_general(cursor, record->ip,
1084                                               &record->leaf);
1085                 if (error && error != ENOENT)
1086                         goto done;
1087         }
1088
1089         /*
1090          * Setup the cursor.
1091          */
1092         hammer_normalize_cursor(cursor);
1093         cursor->key_beg = record->leaf.base;
1094         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1095         cursor->flags |= HAMMER_CURSOR_BACKEND;
1096         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1097
1098         /*
1099          * Records can wind up on-media before the inode itself is on-media.
1100          * Flag the case.
1101          */
1102         record->ip->flags |= HAMMER_INODE_DONDISK;
1103
1104         /*
1105          * If we are deleting a directory entry an exact match must be
1106          * found on-disk.
1107          */
1108         if (record->type == HAMMER_MEM_RECORD_DEL) {
1109                 error = hammer_btree_lookup(cursor);
1110                 if (error == 0) {
1111                         KKASSERT(cursor->iprec == NULL);
1112                         error = hammer_ip_delete_record(cursor, record->ip,
1113                                                         trans->tid);
1114                         if (error == 0) {
1115                                 record->flags |= HAMMER_RECF_DELETED_FE;
1116                                 record->flags |= HAMMER_RECF_DELETED_BE;
1117                                 record->flags |= HAMMER_RECF_COMMITTED;
1118                         }
1119                 }
1120                 goto done;
1121         }
1122
1123         /*
1124          * We are inserting.
1125          *
1126          * Issue a lookup to position the cursor and locate the cluster.  The
1127          * target key should not exist.  If we are creating a directory entry
1128          * we may have to iterate the low 32 bits of the key to find an unused
1129          * key.
1130          */
1131         hammer_sync_lock_sh(trans);
1132         cursor->flags |= HAMMER_CURSOR_INSERT;
1133         error = hammer_btree_lookup(cursor);
1134         if (hammer_debug_inode)
1135                 kprintf("DOINSERT LOOKUP %d\n", error);
1136         if (error == 0) {
1137                 kprintf("hammer_ip_sync_record: duplicate rec "
1138                         "at (%016llx)\n", record->leaf.base.key);
1139                 Debugger("duplicate record1");
1140                 error = EIO;
1141         }
1142 #if 0
1143         if (record->type == HAMMER_MEM_RECORD_DATA)
1144                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1145                         record->leaf.base.key - record->leaf.data_len,
1146                         record->leaf.data_offset, error);
1147 #endif
1148
1149         if (error != ENOENT)
1150                 goto done_unlock;
1151
1152         /*
1153          * Allocate the record and data.  The result buffers will be
1154          * marked as being modified and further calls to
1155          * hammer_modify_buffer() will result in unneeded UNDO records.
1156          *
1157          * Support zero-fill records (data == NULL and data_len != 0)
1158          */
1159         if (record->type == HAMMER_MEM_RECORD_DATA) {
1160                 /*
1161                  * The data portion of a bulk-data record has already been
1162                  * committed to disk, we need only adjust the layer2
1163                  * statistics in the same transaction as our B-Tree insert.
1164                  */
1165                 KKASSERT(record->leaf.data_offset != 0);
1166                 error = hammer_blockmap_finalize(trans,
1167                                                  record->resv,
1168                                                  record->leaf.data_offset,
1169                                                  record->leaf.data_len);
1170         } else if (record->data && record->leaf.data_len) {
1171                 /*
1172                  * Wholely cached record, with data.  Allocate the data.
1173                  */
1174                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1175                                           record->leaf.base.rec_type,
1176                                           &record->leaf.data_offset,
1177                                           &cursor->data_buffer, &error);
1178                 if (bdata == NULL)
1179                         goto done_unlock;
1180                 hammer_crc_set_leaf(record->data, &record->leaf);
1181                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1182                 bcopy(record->data, bdata, record->leaf.data_len);
1183                 hammer_modify_buffer_done(cursor->data_buffer);
1184         } else {
1185                 /*
1186                  * Wholely cached record, without data.
1187                  */
1188                 record->leaf.data_offset = 0;
1189                 record->leaf.data_crc = 0;
1190         }
1191
1192         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1193         if (hammer_debug_inode && error)
1194                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1195
1196         /*
1197          * Our record is on-disk, normally mark the in-memory version as
1198          * deleted.  If the record represented a directory deletion but
1199          * we had to sync a valid directory entry to disk we must convert
1200          * the record to a covering delete so the frontend does not have
1201          * visibility on the synced entry.
1202          */
1203         if (error == 0) {
1204                 if (doprop) {
1205                         hammer_btree_do_propagation(cursor,
1206                                                     record->ip->pfsm,
1207                                                     &record->leaf);
1208                 }
1209                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1210                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1211                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1212                         record->type = HAMMER_MEM_RECORD_DEL;
1213                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1214                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1215                         /* hammer_flush_record_done takes care of the rest */
1216                 } else {
1217                         record->flags |= HAMMER_RECF_DELETED_FE;
1218                         record->flags |= HAMMER_RECF_DELETED_BE;
1219                 }
1220                 record->flags |= HAMMER_RECF_COMMITTED;
1221         } else {
1222                 if (record->leaf.data_offset) {
1223                         hammer_blockmap_free(trans, record->leaf.data_offset,
1224                                              record->leaf.data_len);
1225                 }
1226         }
1227 done_unlock:
1228         hammer_sync_unlock(trans);
1229 done:
1230         return(error);
1231 }
1232
1233 /*
1234  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1235  * entry's key is used to deal with hash collisions in the upper 32 bits.
1236  * A unique 64 bit key is generated in-memory and may be regenerated a
1237  * second time when the directory record is flushed to the on-disk B-Tree.
1238  *
1239  * A referenced record is passed to this function.  This function
1240  * eats the reference.  If an error occurs the record will be deleted.
1241  *
1242  * A copy of the temporary record->data pointer provided by the caller
1243  * will be made.
1244  */
1245 int
1246 hammer_mem_add(hammer_record_t record)
1247 {
1248         hammer_mount_t hmp = record->ip->hmp;
1249
1250         /*
1251          * Make a private copy of record->data
1252          */
1253         if (record->data)
1254                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1255
1256         /*
1257          * Insert into the RB tree.  A unique key should have already
1258          * been selected if this is a directory entry.
1259          */
1260         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1261                 record->flags |= HAMMER_RECF_DELETED_FE;
1262                 hammer_rel_mem_record(record);
1263                 return (EEXIST);
1264         }
1265         ++hmp->count_newrecords;
1266         ++hmp->rsv_recs;
1267         ++record->ip->rsv_recs;
1268         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1269         record->flags |= HAMMER_RECF_ONRBTREE;
1270         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1271         hammer_rel_mem_record(record);
1272         return(0);
1273 }
1274
1275 /************************************************************************
1276  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1277  ************************************************************************
1278  *
1279  * These functions augment the B-Tree scanning functions in hammer_btree.c
1280  * by merging in-memory records with on-disk records.
1281  */
1282
1283 /*
1284  * Locate a particular record either in-memory or on-disk.
1285  *
1286  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1287  * NOT be called to iterate results.
1288  */
1289 int
1290 hammer_ip_lookup(hammer_cursor_t cursor)
1291 {
1292         int error;
1293
1294         /*
1295          * If the element is in-memory return it without searching the
1296          * on-disk B-Tree
1297          */
1298         KKASSERT(cursor->ip);
1299         error = hammer_mem_lookup(cursor);
1300         if (error == 0) {
1301                 cursor->leaf = &cursor->iprec->leaf;
1302                 return(error);
1303         }
1304         if (error != ENOENT)
1305                 return(error);
1306
1307         /*
1308          * If the inode has on-disk components search the on-disk B-Tree.
1309          */
1310         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1311                 return(error);
1312         error = hammer_btree_lookup(cursor);
1313         if (error == 0)
1314                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1315         return(error);
1316 }
1317
1318 /*
1319  * Locate the first record within the cursor's key_beg/key_end range,
1320  * restricted to a particular inode.  0 is returned on success, ENOENT
1321  * if no records matched the requested range, or some other error.
1322  *
1323  * When 0 is returned hammer_ip_next() may be used to iterate additional
1324  * records within the requested range.
1325  *
1326  * This function can return EDEADLK, requiring the caller to terminate
1327  * the cursor and try again.
1328  */
1329 int
1330 hammer_ip_first(hammer_cursor_t cursor)
1331 {
1332         hammer_inode_t ip = cursor->ip;
1333         int error;
1334
1335         KKASSERT(ip != NULL);
1336
1337         /*
1338          * Clean up fields and setup for merged scan
1339          */
1340         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1341         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1342         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1343         if (cursor->iprec) {
1344                 hammer_rel_mem_record(cursor->iprec);
1345                 cursor->iprec = NULL;
1346         }
1347
1348         /*
1349          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1350          * exact lookup so if we get ENOENT we have to call the iterate
1351          * function to validate the first record after the begin key.
1352          *
1353          * The ATEDISK flag is used by hammer_btree_iterate to determine
1354          * whether it must index forwards or not.  It is also used here
1355          * to select the next record from in-memory or on-disk.
1356          *
1357          * EDEADLK can only occur if the lookup hit an empty internal
1358          * element and couldn't delete it.  Since this could only occur
1359          * in-range, we can just iterate from the failure point.
1360          */
1361         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1362                 error = hammer_btree_lookup(cursor);
1363                 if (error == ENOENT || error == EDEADLK) {
1364                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1365                         if (hammer_debug_general & 0x2000)
1366                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1367                         error = hammer_btree_iterate(cursor);
1368                 }
1369                 if (error && error != ENOENT)
1370                         return(error);
1371                 if (error == 0) {
1372                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1373                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1374                 } else {
1375                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1376                 }
1377         }
1378
1379         /*
1380          * Search the in-memory record list (Red-Black tree).  Unlike the
1381          * B-Tree search, mem_first checks for records in the range.
1382          */
1383         error = hammer_mem_first(cursor);
1384         if (error && error != ENOENT)
1385                 return(error);
1386         if (error == 0) {
1387                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1388                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1389                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1390                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1391         }
1392
1393         /*
1394          * This will return the first matching record.
1395          */
1396         return(hammer_ip_next(cursor));
1397 }
1398
1399 /*
1400  * Retrieve the next record in a merged iteration within the bounds of the
1401  * cursor.  This call may be made multiple times after the cursor has been
1402  * initially searched with hammer_ip_first().
1403  *
1404  * 0 is returned on success, ENOENT if no further records match the
1405  * requested range, or some other error code is returned.
1406  */
1407 int
1408 hammer_ip_next(hammer_cursor_t cursor)
1409 {
1410         hammer_btree_elm_t elm;
1411         hammer_record_t rec, save;
1412         int error;
1413         int r;
1414
1415 next_btree:
1416         /*
1417          * Load the current on-disk and in-memory record.  If we ate any
1418          * records we have to get the next one.
1419          *
1420          * If we deleted the last on-disk record we had scanned ATEDISK will
1421          * be clear and RETEST will be set, forcing a call to iterate.  The
1422          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1423          * element.  If ATEDISK is set, iterate will skip the 'current'
1424          * element.
1425          *
1426          * Get the next on-disk record
1427          */
1428         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_RETEST)) {
1429                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1430                         error = hammer_btree_iterate(cursor);
1431                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1432                         if (error == 0) {
1433                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1434                                 hammer_cache_node(&cursor->ip->cache[1],
1435                                                   cursor->node);
1436                         } else {
1437                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1438                                                  HAMMER_CURSOR_ATEDISK;
1439                         }
1440                 }
1441         }
1442
1443 next_memory:
1444         /*
1445          * Get the next in-memory record.
1446          *
1447          * hammer_rec_scan_cmp:  Is the record still in our general range,
1448          *                       (non-inclusive of snapshot exclusions)?
1449          * hammer_rec_scan_callback: Is the record in our snapshot?
1450          */
1451         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1452                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1453                         save = cursor->iprec;
1454                         cursor->iprec = NULL;
1455                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1456                         while (rec) {
1457                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1458                                         break;
1459                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1460                                         break;
1461                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1462                         }
1463                         if (save)
1464                                 hammer_rel_mem_record(save);
1465                         if (cursor->iprec) {
1466                                 KKASSERT(cursor->iprec == rec);
1467                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1468                         } else {
1469                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1470                         }
1471                 }
1472         }
1473
1474         /*
1475          * The memory record may have become stale while being held in
1476          * cursor->iprec.  We are interlocked against the backend on
1477          * with regards to B-Tree entries.
1478          */
1479         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1480                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1481                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1482                         goto next_memory;
1483                 }
1484         }
1485
1486         /*
1487          * Extract either the disk or memory record depending on their
1488          * relative position.
1489          */
1490         error = 0;
1491         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1492         case 0:
1493                 /*
1494                  * Both entries valid.   Compare the entries and nominally
1495                  * return the first one in the sort order.  Numerous cases
1496                  * require special attention, however.
1497                  */
1498                 elm = &cursor->node->ondisk->elms[cursor->index];
1499                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1500
1501                 /*
1502                  * If the two entries differ only by their key (-2/2) or
1503                  * create_tid (-1/1), and are DATA records, we may have a
1504                  * nominal match.  We have to calculate the base file
1505                  * offset of the data.
1506                  */
1507                 if (r <= 2 && r >= -2 && r != 0 &&
1508                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1509                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1510                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1511                         int64_t base2 = cursor->iprec->leaf.base.key -
1512                                         cursor->iprec->leaf.data_len;
1513                         if (base1 == base2)
1514                                 r = 0;
1515                 }
1516
1517                 if (r < 0) {
1518                         error = hammer_btree_extract(cursor,
1519                                                      HAMMER_CURSOR_GET_LEAF);
1520                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1521                         break;
1522                 }
1523
1524                 /*
1525                  * If the entries match exactly the memory entry is either
1526                  * an on-disk directory entry deletion or a bulk data
1527                  * overwrite.  If it is a directory entry deletion we eat
1528                  * both entries.
1529                  *
1530                  * For the bulk-data overwrite case it is possible to have
1531                  * visibility into both, which simply means the syncer
1532                  * hasn't gotten around to doing the delete+insert sequence
1533                  * on the B-Tree.  Use the memory entry and throw away the
1534                  * on-disk entry.
1535                  *
1536                  * If the in-memory record is not either of these we
1537                  * probably caught the syncer while it was syncing it to
1538                  * the media.  Since we hold a shared lock on the cursor,
1539                  * the in-memory record had better be marked deleted at
1540                  * this point.
1541                  */
1542                 if (r == 0) {
1543                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1544                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1545                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1546                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1547                                         goto next_btree;
1548                                 }
1549                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1550                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1551                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1552                                 }
1553                                 /* fall through to memory entry */
1554                         } else {
1555                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1556                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1557                                 goto next_memory;
1558                         }
1559                 }
1560                 /* fall through to the memory entry */
1561         case HAMMER_CURSOR_ATEDISK:
1562                 /*
1563                  * Only the memory entry is valid.
1564                  */
1565                 cursor->leaf = &cursor->iprec->leaf;
1566                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1567
1568                 /*
1569                  * If the memory entry is an on-disk deletion we should have
1570                  * also had found a B-Tree record.  If the backend beat us
1571                  * to it it would have interlocked the cursor and we should
1572                  * have seen the in-memory record marked DELETED_FE.
1573                  */
1574                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1575                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1576                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1577                 }
1578                 break;
1579         case HAMMER_CURSOR_ATEMEM:
1580                 /*
1581                  * Only the disk entry is valid
1582                  */
1583                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1584                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1585                 break;
1586         default:
1587                 /*
1588                  * Neither entry is valid
1589                  *
1590                  * XXX error not set properly
1591                  */
1592                 cursor->leaf = NULL;
1593                 error = ENOENT;
1594                 break;
1595         }
1596         return(error);
1597 }
1598
1599 /*
1600  * Resolve the cursor->data pointer for the current cursor position in
1601  * a merged iteration.
1602  */
1603 int
1604 hammer_ip_resolve_data(hammer_cursor_t cursor)
1605 {
1606         hammer_record_t record;
1607         int error;
1608
1609         if (hammer_cursor_inmem(cursor)) {
1610                 /*
1611                  * The data associated with an in-memory record is usually
1612                  * kmalloced, but reserve-ahead data records will have an
1613                  * on-disk reference.
1614                  *
1615                  * NOTE: Reserve-ahead data records must be handled in the
1616                  * context of the related high level buffer cache buffer
1617                  * to interlock against async writes.
1618                  */
1619                 record = cursor->iprec;
1620                 cursor->data = record->data;
1621                 error = 0;
1622                 if (cursor->data == NULL) {
1623                         KKASSERT(record->leaf.base.rec_type ==
1624                                  HAMMER_RECTYPE_DATA);
1625                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1626                                                     record->leaf.data_offset,
1627                                                     record->leaf.data_len,
1628                                                     &error,
1629                                                     &cursor->data_buffer);
1630                 }
1631         } else {
1632                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1633                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1634         }
1635         return(error);
1636 }
1637
1638 /*
1639  * Backend truncation / record replacement - delete records in range.
1640  *
1641  * Delete all records within the specified range for inode ip.  In-memory
1642  * records still associated with the frontend are ignored.
1643  *
1644  * If truncating is non-zero in-memory records associated with the back-end
1645  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1646  *
1647  * NOTES:
1648  *
1649  *      * An unaligned range will cause new records to be added to cover
1650  *        the edge cases. (XXX not implemented yet).
1651  *
1652  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1653  *        also do not deal with unaligned ranges.
1654  *
1655  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1656  *
1657  *      * Record keys for regular file data have to be special-cased since
1658  *        they indicate the end of the range (key = base + bytes).
1659  *
1660  *      * This function may be asked to delete ridiculously huge ranges, for
1661  *        example if someone truncates or removes a 1TB regular file.  We
1662  *        must be very careful on restarts and we may have to stop w/
1663  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1664  */
1665 int
1666 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1667                        int64_t ran_beg, int64_t ran_end, int truncating)
1668 {
1669         hammer_transaction_t trans = cursor->trans;
1670         hammer_btree_leaf_elm_t leaf;
1671         int error;
1672         int64_t off;
1673         int64_t tmp64;
1674
1675 #if 0
1676         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1677 #endif
1678
1679         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1680 retry:
1681         hammer_normalize_cursor(cursor);
1682         cursor->key_beg.localization = ip->obj_localization +
1683                                        HAMMER_LOCALIZE_MISC;
1684         cursor->key_beg.obj_id = ip->obj_id;
1685         cursor->key_beg.create_tid = 0;
1686         cursor->key_beg.delete_tid = 0;
1687         cursor->key_beg.obj_type = 0;
1688
1689         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1690                 cursor->key_beg.key = ran_beg;
1691                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1692         } else {
1693                 /*
1694                  * The key in the B-Tree is (base+bytes), so the first possible
1695                  * matching key is ran_beg + 1.
1696                  */
1697                 cursor->key_beg.key = ran_beg + 1;
1698                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1699         }
1700
1701         cursor->key_end = cursor->key_beg;
1702         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1703                 cursor->key_end.key = ran_end;
1704         } else {
1705                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1706                 if (tmp64 < ran_end)
1707                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1708                 else
1709                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1710         }
1711
1712         cursor->asof = ip->obj_asof;
1713         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1714         cursor->flags |= HAMMER_CURSOR_ASOF;
1715         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1716         cursor->flags |= HAMMER_CURSOR_BACKEND;
1717         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1718
1719         error = hammer_ip_first(cursor);
1720
1721         /*
1722          * Iterate through matching records and mark them as deleted.
1723          */
1724         while (error == 0) {
1725                 leaf = cursor->leaf;
1726
1727                 KKASSERT(leaf->base.delete_tid == 0);
1728                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1729
1730                 /*
1731                  * There may be overlap cases for regular file data.  Also
1732                  * remember the key for a regular file record is (base + len),
1733                  * NOT (base).
1734                  *
1735                  * Note that do to duplicates (mem & media) allowed by
1736                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1737                  */
1738                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1739                         off = leaf->base.key - leaf->data_len;
1740                         /*
1741                          * Check the left edge case.  We currently do not
1742                          * split existing records.
1743                          */
1744                         if (off < ran_beg && leaf->base.key > ran_beg) {
1745                                 panic("hammer left edge case %016llx %d\n",
1746                                         leaf->base.key, leaf->data_len);
1747                         }
1748
1749                         /*
1750                          * Check the right edge case.  Note that the
1751                          * record can be completely out of bounds, which
1752                          * terminates the search.
1753                          *
1754                          * base->key is exclusive of the right edge while
1755                          * ran_end is inclusive of the right edge.  The
1756                          * (key - data_len) left boundary is inclusive.
1757                          *
1758                          * XXX theory-check this test at some point, are
1759                          * we missing a + 1 somewhere?  Note that ran_end
1760                          * could overflow.
1761                          */
1762                         if (leaf->base.key - 1 > ran_end) {
1763                                 if (leaf->base.key - leaf->data_len > ran_end)
1764                                         break;
1765                                 panic("hammer right edge case\n");
1766                         }
1767                 } else {
1768                         off = leaf->base.key;
1769                 }
1770
1771                 /*
1772                  * Delete the record.  When truncating we do not delete
1773                  * in-memory (data) records because they represent data
1774                  * written after the truncation.
1775                  *
1776                  * This will also physically destroy the B-Tree entry and
1777                  * data if the retention policy dictates.  The function
1778                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
1779                  * to retest the new 'current' element.
1780                  */
1781                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1782                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1783                         /*
1784                          * If we have built up too many meta-buffers we risk
1785                          * deadlocking the kernel and must stop.  This can
1786                          * occur when deleting ridiculously huge files.
1787                          * sync_trunc_off is updated so the next cycle does
1788                          * not re-iterate records we have already deleted.
1789                          *
1790                          * This is only done with formal truncations.
1791                          */
1792                         if (truncating > 1 && error == 0 &&
1793                             hammer_flusher_meta_limit(ip->hmp)) {
1794                                 ip->sync_trunc_off = off;
1795                                 error = EWOULDBLOCK;
1796                         }
1797                 }
1798                 if (error)
1799                         break;
1800                 ran_beg = off;  /* for restart */
1801                 error = hammer_ip_next(cursor);
1802         }
1803         if (cursor->node)
1804                 hammer_cache_node(&ip->cache[1], cursor->node);
1805
1806         if (error == EDEADLK) {
1807                 hammer_done_cursor(cursor);
1808                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1809                 if (error == 0)
1810                         goto retry;
1811         }
1812         if (error == ENOENT)
1813                 error = 0;
1814         return(error);
1815 }
1816
1817 /*
1818  * This backend function deletes the specified record on-disk, similar to
1819  * delete_range but for a specific record.  Unlike the exact deletions
1820  * used when deleting a directory entry this function uses an ASOF search
1821  * like delete_range.
1822  *
1823  * This function may be called with ip->obj_asof set for a slave snapshot,
1824  * so don't use it.  We always delete non-historical records only.
1825  */
1826 static int
1827 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
1828                       hammer_btree_leaf_elm_t leaf)
1829 {
1830         hammer_transaction_t trans = cursor->trans;
1831         int error;
1832
1833         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1834 retry:
1835         hammer_normalize_cursor(cursor);
1836         cursor->key_beg = leaf->base;
1837         cursor->asof = HAMMER_MAX_TID;
1838         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1839         cursor->flags |= HAMMER_CURSOR_ASOF;
1840         cursor->flags |= HAMMER_CURSOR_BACKEND;
1841         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1842
1843         error = hammer_btree_lookup(cursor);
1844         if (error == 0) {
1845                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1846         }
1847         if (error == EDEADLK) {
1848                 hammer_done_cursor(cursor);
1849                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1850                 if (error == 0)
1851                         goto retry;
1852         }
1853         return(error);
1854 }
1855
1856 /*
1857  * This function deletes remaining auxillary records when an inode is
1858  * being deleted.  This function explicitly does not delete the
1859  * inode record, directory entry, data, or db records.  Those must be
1860  * properly disposed of prior to this call.
1861  */
1862 int
1863 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
1864 {
1865         hammer_transaction_t trans = cursor->trans;
1866         hammer_btree_leaf_elm_t leaf;
1867         int error;
1868
1869         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1870 retry:
1871         hammer_normalize_cursor(cursor);
1872         cursor->key_beg.localization = ip->obj_localization +
1873                                        HAMMER_LOCALIZE_MISC;
1874         cursor->key_beg.obj_id = ip->obj_id;
1875         cursor->key_beg.create_tid = 0;
1876         cursor->key_beg.delete_tid = 0;
1877         cursor->key_beg.obj_type = 0;
1878         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
1879         cursor->key_beg.key = HAMMER_MIN_KEY;
1880
1881         cursor->key_end = cursor->key_beg;
1882         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
1883         cursor->key_end.key = HAMMER_MAX_KEY;
1884
1885         cursor->asof = ip->obj_asof;
1886         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1887         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1888         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1889         cursor->flags |= HAMMER_CURSOR_BACKEND;
1890
1891         error = hammer_ip_first(cursor);
1892
1893         /*
1894          * Iterate through matching records and mark them as deleted.
1895          */
1896         while (error == 0) {
1897                 leaf = cursor->leaf;
1898
1899                 KKASSERT(leaf->base.delete_tid == 0);
1900
1901                 /*
1902                  * Mark the record and B-Tree entry as deleted.  This will
1903                  * also physically delete the B-Tree entry, record, and
1904                  * data if the retention policy dictates.  The function
1905                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
1906                  * to retest the new 'current' element.
1907                  *
1908                  * Directory entries (and delete-on-disk directory entries)
1909                  * must be synced and cannot be deleted.
1910                  */
1911                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1912                 ++*countp;
1913                 if (error)
1914                         break;
1915                 error = hammer_ip_next(cursor);
1916         }
1917         if (cursor->node)
1918                 hammer_cache_node(&ip->cache[1], cursor->node);
1919         if (error == EDEADLK) {
1920                 hammer_done_cursor(cursor);
1921                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1922                 if (error == 0)
1923                         goto retry;
1924         }
1925         if (error == ENOENT)
1926                 error = 0;
1927         return(error);
1928 }
1929
1930 /*
1931  * Delete the record at the current cursor.  On success the cursor will
1932  * be positioned appropriately for an iteration but may no longer be at
1933  * a leaf node.
1934  *
1935  * This routine is only called from the backend.
1936  *
1937  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1938  * cursor and retry.
1939  */
1940 int
1941 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1942                         hammer_tid_t tid)
1943 {
1944         hammer_record_t iprec;
1945         hammer_mount_t hmp;
1946         int error;
1947
1948         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1949         KKASSERT(tid != 0);
1950         hmp = cursor->node->hmp;
1951
1952         /*
1953          * In-memory (unsynchronized) records can simply be freed.  This
1954          * only occurs in range iterations since all other records are
1955          * individually synchronized.  Thus there should be no confusion with
1956          * the interlock.
1957          *
1958          * An in-memory record may be deleted before being committed to disk,
1959          * but could have been accessed in the mean time.  The reservation
1960          * code will deal with the case.
1961          */
1962         if (hammer_cursor_inmem(cursor)) {
1963                 iprec = cursor->iprec;
1964                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1965                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1966                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1967                 return(0);
1968         }
1969
1970         /*
1971          * On-disk records are marked as deleted by updating their delete_tid.
1972          * This does not effect their position in the B-Tree (which is based
1973          * on their create_tid).
1974          *
1975          * Frontend B-Tree operations track inodes so we tell
1976          * hammer_delete_at_cursor() not to.
1977          */
1978         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1979
1980         if (error == 0) {
1981                 error = hammer_delete_at_cursor(
1982                                 cursor,
1983                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
1984                                 cursor->trans->tid,
1985                                 cursor->trans->time32,
1986                                 0, NULL);
1987         }
1988         return(error);
1989 }
1990
1991 /*
1992  * Delete the B-Tree element at the current cursor and do any necessary
1993  * mirror propagation.
1994  *
1995  * The cursor must be properly positioned for an iteration on return but
1996  * may be pointing at an internal element.
1997  *
1998  * An element can be un-deleted by passing a delete_tid of 0 with
1999  * HAMMER_DELETE_ADJUST.
2000  */
2001 int
2002 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2003                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2004                         int track, int64_t *stat_bytes)
2005 {
2006         struct hammer_btree_leaf_elm save_leaf;
2007         hammer_transaction_t trans;
2008         hammer_btree_leaf_elm_t leaf;
2009         hammer_node_t node;
2010         hammer_btree_elm_t elm;
2011         hammer_off_t data_offset;
2012         int32_t data_len;
2013         u_int16_t rec_type;
2014         int error;
2015         int icount;
2016         int doprop;
2017
2018         error = hammer_cursor_upgrade(cursor);
2019         if (error)
2020                 return(error);
2021
2022         trans = cursor->trans;
2023         node = cursor->node;
2024         elm = &node->ondisk->elms[cursor->index];
2025         leaf = &elm->leaf;
2026         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2027
2028         hammer_sync_lock_sh(trans);
2029         doprop = 0;
2030         icount = 0;
2031
2032         /*
2033          * Adjust the delete_tid.  Update the mirror_tid propagation field
2034          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2035          */
2036         if (delete_flags & HAMMER_DELETE_ADJUST) {
2037                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2038                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2039                                 icount = -1;
2040                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2041                                 icount = 1;
2042                 }
2043
2044                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2045                 elm->leaf.base.delete_tid = delete_tid;
2046                 elm->leaf.delete_ts = delete_ts;
2047                 hammer_modify_node_done(node);
2048
2049                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2050                         hammer_modify_node_field(trans, node, mirror_tid);
2051                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2052                         hammer_modify_node_done(node);
2053                         doprop = 1;
2054                         if (hammer_debug_general & 0x0002) {
2055                                 kprintf("delete_at_cursor: propagate %016llx"
2056                                         " @%016llx\n",
2057                                         elm->leaf.base.delete_tid,
2058                                         node->node_offset);
2059                         }
2060                 }
2061
2062                 /*
2063                  * Adjust for the iteration.  We have deleted the current
2064                  * element and want to clear ATEDISK so the iteration does
2065                  * not skip the element after, which now becomes the current
2066                  * element.  This element must be re-tested if doing an
2067                  * iteration, which is handled by the RETEST flag.
2068                  */
2069                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2070                         cursor->flags |= HAMMER_CURSOR_RETEST;
2071                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2072                 }
2073
2074                 /*
2075                  * An on-disk record cannot have the same delete_tid
2076                  * as its create_tid.  In a chain of record updates
2077                  * this could result in a duplicate record.
2078                  */
2079                 KKASSERT(elm->leaf.base.delete_tid !=
2080                          elm->leaf.base.create_tid);
2081         }
2082
2083         /*
2084          * Destroy the B-Tree element if asked (typically if a nohistory
2085          * file or mount, or when called by the pruning code).
2086          *
2087          * Adjust the ATEDISK flag to properly support iterations.
2088          */
2089         if (delete_flags & HAMMER_DELETE_DESTROY) {
2090                 data_offset = elm->leaf.data_offset;
2091                 data_len = elm->leaf.data_len;
2092                 rec_type = elm->leaf.base.rec_type;
2093                 if (doprop) {
2094                         save_leaf = elm->leaf;
2095                         leaf = &save_leaf;
2096                 }
2097                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2098                     elm->leaf.base.delete_tid == 0) {
2099                         icount = -1;
2100                 }
2101
2102                 error = hammer_btree_delete(cursor);
2103                 if (error == 0) {
2104                         /*
2105                          * The deletion moves the next element (if any) to
2106                          * the current element position.  We must clear
2107                          * ATEDISK so this element is not skipped and we
2108                          * must set RETEST to force any iteration to re-test
2109                          * the element.
2110                          */
2111                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2112                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2113                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2114                         }
2115                 }
2116                 if (error == 0) {
2117                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2118                         case HAMMER_ZONE_LARGE_DATA:
2119                         case HAMMER_ZONE_SMALL_DATA:
2120                         case HAMMER_ZONE_META:
2121                                 hammer_blockmap_free(trans,
2122                                                      data_offset, data_len);
2123                                 break;
2124                         default:
2125                                 break;
2126                         }
2127                 }
2128         }
2129
2130         /*
2131          * Track inode count and next_tid.  This is used by the mirroring
2132          * and PFS code.  icount can be negative, zero, or positive.
2133          */
2134         if (error == 0 && track) {
2135                 if (icount) {
2136                         hammer_modify_volume_field(trans, trans->rootvol,
2137                                                    vol0_stat_inodes);
2138                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2139                         hammer_modify_volume_done(trans->rootvol);
2140                 }
2141                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2142                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2143                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2144                         hammer_modify_volume_done(trans->rootvol);
2145                 }
2146         }
2147
2148         /*
2149          * mirror_tid propagation occurs if the node's mirror_tid had to be
2150          * updated while adjusting the delete_tid.
2151          *
2152          * This occurs when deleting even in nohistory mode, but does not
2153          * occur when pruning an already-deleted node.
2154          *
2155          * cursor->ip is NULL when called from the pruning, mirroring,
2156          * and pfs code.  If non-NULL propagation will be conditionalized
2157          * on whether the PFS is in no-history mode or not.
2158          */
2159         if (doprop) {
2160                 if (cursor->ip)
2161                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2162                 else
2163                         hammer_btree_do_propagation(cursor, NULL, leaf);
2164         }
2165         hammer_sync_unlock(trans);
2166         return (error);
2167 }
2168
2169 /*
2170  * Determine whether we can remove a directory.  This routine checks whether
2171  * a directory is empty or not and enforces flush connectivity.
2172  *
2173  * Flush connectivity requires that we block if the target directory is
2174  * currently flushing, otherwise it may not end up in the same flush group.
2175  *
2176  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2177  */
2178 int
2179 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2180 {
2181         struct hammer_cursor cursor;
2182         int error;
2183
2184         /*
2185          * Check directory empty
2186          */
2187         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2188
2189         cursor.key_beg.localization = ip->obj_localization +
2190                                       HAMMER_LOCALIZE_MISC;
2191         cursor.key_beg.obj_id = ip->obj_id;
2192         cursor.key_beg.create_tid = 0;
2193         cursor.key_beg.delete_tid = 0;
2194         cursor.key_beg.obj_type = 0;
2195         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2196         cursor.key_beg.key = HAMMER_MIN_KEY;
2197
2198         cursor.key_end = cursor.key_beg;
2199         cursor.key_end.rec_type = 0xFFFF;
2200         cursor.key_end.key = HAMMER_MAX_KEY;
2201
2202         cursor.asof = ip->obj_asof;
2203         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2204
2205         error = hammer_ip_first(&cursor);
2206         if (error == ENOENT)
2207                 error = 0;
2208         else if (error == 0)
2209                 error = ENOTEMPTY;
2210         hammer_done_cursor(&cursor);
2211         return(error);
2212 }
2213