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 1.89 2008/07/13 09:32:48 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_add(hammer_record_t record);
  40 static int hammer_mem_lookup(hammer_cursor_t cursor);
  41 static int hammer_mem_first(hammer_cursor_t cursor);
  42 static int hammer_frontend_trunc_callback(hammer_record_t record,
  43                                 void *data __unused);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45 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 /*
  54  * Red-black tree support.  Comparison code for insertion.
  55  */
  56 static int
  57 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  58 {
  59         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  60                 return(-1);
  61         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  62                 return(1);
  63
  64         if (rec1->leaf.base.key < rec2->leaf.base.key)
  65                 return(-1);
  66         if (rec1->leaf.base.key > rec2->leaf.base.key)
  67                 return(1);
  68
  69         /*
  70          * Never match against an item deleted by the front-end.
  71          *
  72          * rec1 is greater then rec2 if rec1 is marked deleted.
  73          * rec1 is less then rec2 if rec2 is marked deleted.
  74          *
  75          * Multiple deleted records may be present, do not return 0
  76          * if both are marked deleted.
  77          */
  78         if (rec1->flags & HAMMER_RECF_DELETED_FE)
  79                 return(1);
  80         if (rec2->flags & HAMMER_RECF_DELETED_FE)
  81                 return(-1);
  82
  83         return(0);
  84 }
  85
  86 /*
  87  * Basic record comparison code similar to hammer_btree_cmp().
  88  */
  89 static int
  90 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
  91 {
  92         if (elm->rec_type < rec->leaf.base.rec_type)
  93                 return(-3);
  94         if (elm->rec_type > rec->leaf.base.rec_type)
  95                 return(3);
  96
  97         if (elm->key < rec->leaf.base.key)
  98                 return(-2);
  99         if (elm->key > rec->leaf.base.key)
 100                 return(2);
 101
 102         /*
 103          * Never match against an item deleted by the front-end.
 104          * elm is less then rec if rec is marked deleted.
 105          */
 106         if (rec->flags & HAMMER_RECF_DELETED_FE)
 107                 return(-1);
 108         return(0);
 109 }
 110
 111 /*
 112  * Special LOOKUP_INFO to locate an overlapping record.  This used by
 113  * the reservation code to implement small-block records (whos keys will
 114  * be different depending on data_len, when representing the same base
 115  * offset).
 116  *
 117  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 118  */
 119 static int
 120 hammer_rec_overlap_compare(hammer_btree_leaf_elm_t leaf, hammer_record_t rec)
 121 {
 122         if (leaf->base.rec_type < rec->leaf.base.rec_type)
 123                 return(-3);
 124         if (leaf->base.rec_type > rec->leaf.base.rec_type)
 125                 return(3);
 126
 127         /*
 128          * Overlap compare
 129          */
 130         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 131                 /* leaf_end <= rec_beg */
 132                 if (leaf->base.key <= rec->leaf.base.key - rec->leaf.data_len)
 133                         return(-2);
 134                 /* leaf_beg >= rec_end */
 135                 if (leaf->base.key - leaf->data_len >= rec->leaf.base.key)
 136                         return(2);
 137         } else {
 138                 if (leaf->base.key < rec->leaf.base.key)
 139                         return(-2);
 140                 if (leaf->base.key > rec->leaf.base.key)
 141                         return(2);
 142         }
 143
 144         /*
 145          * Never match against an item deleted by the front-end.
 146          * leaf is less then rec if rec is marked deleted.
 147          *
 148          * We must still return the proper code for the scan to continue
 149          * along the correct branches.
 150          */
 151         if (rec->flags & HAMMER_RECF_DELETED_FE) {
 152                 if (leaf->base.key < rec->leaf.base.key)
 153                         return(-2);
 154                 if (leaf->base.key > rec->leaf.base.key)
 155                         return(2);
 156                 return(-1);
 157         }
 158         return(0);
 159 }
 160
 161 /*
 162  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 163  * is reversed so the comparison result has to be negated.  key_beg and
 164  * key_end are both range-inclusive.
 165  *
 166  * Localized deletions are not cached in-memory.
 167  */
 168 static
 169 int
 170 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 171 {
 172         hammer_cursor_t cursor = data;
 173         int r;
 174
 175         r = hammer_rec_cmp(&cursor->key_beg, rec);
 176         if (r > 1)
 177                 return(-1);
 178         r = hammer_rec_cmp(&cursor->key_end, rec);
 179         if (r < -1)
 180                 return(1);
 181         return(0);
 182 }
 183
 184 /*
 185  * This compare function is used when simply looking up key_beg.
 186  */
 187 static
 188 int
 189 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 190 {
 191         hammer_cursor_t cursor = data;
 192         int r;
 193
 194         r = hammer_rec_cmp(&cursor->key_beg, rec);
 195         if (r > 1)
 196                 return(-1);
 197         if (r < -1)
 198                 return(1);
 199         return(0);
 200 }
 201
 202 /*
 203  * Locate blocks within the truncation range.  Partial blocks do not count.
 204  */
 205 static
 206 int
 207 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 208 {
 209         struct rec_trunc_info *info = data;
 210
 211         if (rec->leaf.base.rec_type < info->rec_type)
 212                 return(-1);
 213         if (rec->leaf.base.rec_type > info->rec_type)
 214                 return(1);
 215
 216         switch(rec->leaf.base.rec_type) {
 217         case HAMMER_RECTYPE_DB:
 218                 /*
 219                  * DB record key is not beyond the truncation point, retain.
 220                  */
 221                 if (rec->leaf.base.key < info->trunc_off)
 222                         return(-1);
 223                 break;
 224         case HAMMER_RECTYPE_DATA:
 225                 /*
 226                  * DATA record offset start is not beyond the truncation point,
 227                  * retain.
 228                  */
 229                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 230                         return(-1);
 231                 break;
 232         default:
 233                 panic("hammer_rec_trunc_cmp: unexpected record type");
 234         }
 235
 236         /*
 237          * The record start is >= the truncation point, return match,
 238          * the record should be destroyed.
 239          */
 240         return(0);
 241 }
 242
 243 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 244 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 245                     hammer_rec_overlap_compare, hammer_btree_leaf_elm_t);
 246
 247 /*
 248  * Allocate a record for the caller to finish filling in.  The record is
 249  * returned referenced.
 250  */
 251 hammer_record_t
 252 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 253 {
 254         hammer_record_t record;
 255
 256         ++hammer_count_records;
 257         record = kmalloc(sizeof(*record), M_HAMMER,
 258                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 259         record->flush_state = HAMMER_FST_IDLE;
 260         record->ip = ip;
 261         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 262         record->leaf.data_len = data_len;
 263         hammer_ref(&record->lock);
 264
 265         if (data_len) {
 266                 record->data = kmalloc(data_len, M_HAMMER, M_WAITOK | M_ZERO);
 267                 record->flags |= HAMMER_RECF_ALLOCDATA;
 268                 ++hammer_count_record_datas;
 269         }
 270
 271         return (record);
 272 }
 273
 274 void
 275 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 276 {
 277         while (record->flush_state == HAMMER_FST_FLUSH) {
 278                 record->flags |= HAMMER_RECF_WANTED;
 279                 tsleep(record, 0, ident, 0);
 280         }
 281 }
 282
 283 /*
 284  * Called from the backend, hammer_inode.c, after a record has been
 285  * flushed to disk.  The record has been exclusively locked by the
 286  * caller and interlocked with BE.
 287  *
 288  * We clean up the state, unlock, and release the record (the record
 289  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 290  */
 291 void
 292 hammer_flush_record_done(hammer_record_t record, int error)
 293 {
 294         hammer_inode_t target_ip;
 295
 296         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 297         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 298
 299         if (error) {
 300                 /*
 301                  * An error occured, the backend was unable to sync the
 302                  * record to its media.  Leave the record intact.
 303                  */
 304                 Debugger("flush_record_done error");
 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_inode_t ip, target_ip;
 344
 345         hammer_unref(&record->lock);
 346
 347         if (record->lock.refs == 0) {
 348                 /*
 349                  * Upon release of the last reference wakeup any waiters.
 350                  * The record structure may get destroyed so callers will
 351                  * loop up and do a relookup.
 352                  *
 353                  * WARNING!  Record must be removed from RB-TREE before we
 354                  * might possibly block.  hammer_test_inode() can block!
 355                  */
 356                 ip = record->ip;
 357
 358                 /*
 359                  * Upon release of the last reference a record marked deleted
 360                  * is destroyed.
 361                  */
 362                 if (record->flags & HAMMER_RECF_DELETED_FE) {
 363                         KKASSERT(ip->lock.refs > 0);
 364                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 365
 366                         /*
 367                          * target_ip may have zero refs, we have to ref it
 368                          * to prevent it from being ripped out from under
 369                          * us.
 370                          */
 371                         if ((target_ip = record->target_ip) != NULL) {
 372                                 TAILQ_REMOVE(&target_ip->target_list,
 373                                              record, target_entry);
 374                                 record->target_ip = NULL;
 375                                 hammer_ref(&target_ip->lock);
 376                         }
 377
 378                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 379                                 RB_REMOVE(hammer_rec_rb_tree,
 380                                           &record->ip->rec_tree,
 381                                           record);
 382                                 KKASSERT(ip->rsv_recs > 0);
 383                                 --ip->hmp->rsv_recs;
 384                                 --ip->rsv_recs;
 385                                 ip->hmp->rsv_databytes -= record->leaf.data_len;
 386                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 387
 388                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 389                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 390                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 391                                         hammer_test_inode(record->ip);
 392                                 }
 393                         }
 394
 395                         /*
 396                          * Do this test after removing record from the B-Tree.
 397                          */
 398                         if (target_ip) {
 399                                 hammer_test_inode(target_ip);
 400                                 hammer_rel_inode(target_ip, 0);
 401                         }
 402
 403                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 404                                 --hammer_count_record_datas;
 405                                 kfree(record->data, M_HAMMER);
 406                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 407                         }
 408                         if (record->resv) {
 409                                 hammer_blockmap_reserve_complete(ip->hmp,
 410                                                                  record->resv);
 411                                 record->resv = NULL;
 412                         }
 413                         record->data = NULL;
 414                         --hammer_count_records;
 415                         kfree(record, M_HAMMER);
 416                 }
 417         }
 418 }
 419
 420 /*
 421  * Record visibility depends on whether the record is being accessed by
 422  * the backend or the frontend.
 423  *
 424  * Return non-zero if the record is visible, zero if it isn't or if it is
 425  * deleted.
 426  */
 427 static __inline
 428 int
 429 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 430 {
 431         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 432                 if (record->flags & HAMMER_RECF_DELETED_BE)
 433                         return(0);
 434         } else {
 435                 if (record->flags & HAMMER_RECF_DELETED_FE)
 436                         return(0);
 437         }
 438         return(1);
 439 }
 440
 441 /*
 442  * This callback is used as part of the RB_SCAN function for in-memory
 443  * records.  We terminate it (return -1) as soon as we get a match.
 444  *
 445  * This routine is used by frontend code.
 446  *
 447  * The primary compare code does not account for ASOF lookups.  This
 448  * code handles that case as well as a few others.
 449  */
 450 static
 451 int
 452 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 453 {
 454         hammer_cursor_t cursor = data;
 455
 456         /*
 457          * We terminate on success, so this should be NULL on entry.
 458          */
 459         KKASSERT(cursor->iprec == NULL);
 460
 461         /*
 462          * Skip if the record was marked deleted.
 463          */
 464         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 465                 return(0);
 466
 467         /*
 468          * Skip if not visible due to our as-of TID
 469          */
 470         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 471                 if (cursor->asof < rec->leaf.base.create_tid)
 472                         return(0);
 473                 if (rec->leaf.base.delete_tid &&
 474                     cursor->asof >= rec->leaf.base.delete_tid) {
 475                         return(0);
 476                 }
 477         }
 478
 479         /*
 480          * ref the record.  The record is protected from backend B-Tree
 481          * interactions by virtue of the cursor's IP lock.
 482          */
 483         hammer_ref(&rec->lock);
 484
 485         /*
 486          * The record may have been deleted while we were blocked.
 487          */
 488         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 489                 hammer_rel_mem_record(rec);
 490                 return(0);
 491         }
 492
 493         /*
 494          * Set the matching record and stop the scan.
 495          */
 496         cursor->iprec = rec;
 497         return(-1);
 498 }
 499
 500
 501 /*
 502  * Lookup an in-memory record given the key specified in the cursor.  Works
 503  * just like hammer_btree_lookup() but operates on an inode's in-memory
 504  * record list.
 505  *
 506  * The lookup must fail if the record is marked for deferred deletion.
 507  */
 508 static
 509 int
 510 hammer_mem_lookup(hammer_cursor_t cursor)
 511 {
 512         int error;
 513
 514         KKASSERT(cursor->ip);
 515         if (cursor->iprec) {
 516                 hammer_rel_mem_record(cursor->iprec);
 517                 cursor->iprec = NULL;
 518         }
 519         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 520                                    hammer_rec_scan_callback, cursor);
 521
 522         if (cursor->iprec == NULL)
 523                 error = ENOENT;
 524         else
 525                 error = 0;
 526         return(error);
 527 }
 528
 529 /*
 530  * hammer_mem_first() - locate the first in-memory record matching the
 531  * cursor within the bounds of the key range.
 532  */
 533 static
 534 int
 535 hammer_mem_first(hammer_cursor_t cursor)
 536 {
 537         hammer_inode_t ip;
 538
 539         ip = cursor->ip;
 540         KKASSERT(ip != NULL);
 541
 542         if (cursor->iprec) {
 543                 hammer_rel_mem_record(cursor->iprec);
 544                 cursor->iprec = NULL;
 545         }
 546
 547         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 548                                    hammer_rec_scan_callback, cursor);
 549
 550         /*
 551          * Adjust scan.node and keep it linked into the RB-tree so we can
 552          * hold the cursor through third party modifications of the RB-tree.
 553          */
 554         if (cursor->iprec)
 555                 return(0);
 556         return(ENOENT);
 557 }
 558
 559 /************************************************************************
 560  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 561  ************************************************************************
 562  *
 563  * These functions manipulate in-memory records.  Such records typically
 564  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 565  */
 566
 567 /*
 568  * Add a directory entry (dip,ncp) which references inode (ip).
 569  *
 570  * Note that the low 32 bits of the namekey are set temporarily to create
 571  * a unique in-memory record, and may be modified a second time when the
 572  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 573  * all 0's when synching to disk, which is not handled here.
 574  *
 575  * NOTE: bytes does not include any terminating \0 on name, and name might
 576  * not be terminated.
 577  */
 578 int
 579 hammer_ip_add_directory(struct hammer_transaction *trans,
 580                      struct hammer_inode *dip, const char *name, int bytes,
 581                      struct hammer_inode *ip)
 582 {
 583         struct hammer_cursor cursor;
 584         hammer_record_t record;
 585         int error;
 586         int count;
 587         u_int32_t iterator;
 588
 589         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 590         if (++trans->hmp->namekey_iterator == 0)
 591                 ++trans->hmp->namekey_iterator;
 592
 593         record->type = HAMMER_MEM_RECORD_ADD;
 594         record->leaf.base.localization = dip->obj_localization +
 595                                          HAMMER_LOCALIZE_MISC;
 596         record->leaf.base.obj_id = dip->obj_id;
 597         record->leaf.base.key = hammer_directory_namekey(name, bytes);
 598         record->leaf.base.key += trans->hmp->namekey_iterator;
 599         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 600         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 601         record->data->entry.obj_id = ip->obj_id;
 602         record->data->entry.localization = ip->obj_localization;
 603         bcopy(name, record->data->entry.name, bytes);
 604
 605         ++ip->ino_data.nlinks;
 606         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 607
 608         /*
 609          * Find an unused namekey.  Both the in-memory record tree and
 610          * the B-Tree are checked.  Exact matches also match create_tid
 611          * so use an ASOF search to (mostly) ignore it.
 612          *
 613          * delete-visibility is set so pending deletions do not give us
 614          * a false-negative on our ability to use an iterator.
 615          */
 616         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 617         cursor.key_beg = record->leaf.base;
 618         cursor.flags |= HAMMER_CURSOR_ASOF;
 619         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 620         cursor.asof = ip->obj_asof;
 621
 622         count = 0;
 623         while (hammer_ip_lookup(&cursor) == 0) {
 624                 iterator = (u_int32_t)record->leaf.base.key + 1;
 625                 if (iterator == 0)
 626                         iterator = 1;
 627                 record->leaf.base.key &= ~0xFFFFFFFFLL;
 628                 record->leaf.base.key |= iterator;
 629                 cursor.key_beg.key = record->leaf.base.key;
 630                 if (++count == 1000000000) {
 631                         hammer_rel_mem_record(record);
 632                         error = ENOSPC;
 633                         goto failed;
 634                 }
 635         }
 636
 637         /*
 638          * The target inode and the directory entry are bound together.
 639          */
 640         record->target_ip = ip;
 641         record->flush_state = HAMMER_FST_SETUP;
 642         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 643
 644         /*
 645          * The inode now has a dependancy and must be taken out of the idle
 646          * state.  An inode not in an idle state is given an extra reference.
 647          */
 648         if (ip->flush_state == HAMMER_FST_IDLE) {
 649                 hammer_ref(&ip->lock);
 650                 ip->flush_state = HAMMER_FST_SETUP;
 651         }
 652         error = hammer_mem_add(record);
 653 failed:
 654         hammer_done_cursor(&cursor);
 655         return(error);
 656 }
 657
 658 /*
 659  * Delete the directory entry and update the inode link count.  The
 660  * cursor must be seeked to the directory entry record being deleted.
 661  *
 662  * The related inode should be share-locked by the caller.  The caller is
 663  * on the frontend.
 664  *
 665  * This function can return EDEADLK requiring the caller to terminate
 666  * the cursor, any locks, wait on the returned record, and retry.
 667  */
 668 int
 669 hammer_ip_del_directory(struct hammer_transaction *trans,
 670                      hammer_cursor_t cursor, struct hammer_inode *dip,
 671                      struct hammer_inode *ip)
 672 {
 673         hammer_record_t record;
 674         int error;
 675
 676         if (hammer_cursor_inmem(cursor)) {
 677                 /*
 678                  * In-memory (unsynchronized) records can simply be freed.
 679                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 680                  * by the backend, we must still avoid races against the
 681                  * backend potentially syncing the record to the media.
 682                  *
 683                  * We cannot call hammer_ip_delete_record(), that routine may
 684                  * only be called from the backend.
 685                  */
 686                 record = cursor->iprec;
 687                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 688                         KKASSERT(cursor->deadlk_rec == NULL);
 689                         hammer_ref(&record->lock);
 690                         cursor->deadlk_rec = record;
 691                         error = EDEADLK;
 692                 } else {
 693                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 694                         record->flags |= HAMMER_RECF_DELETED_FE;
 695                         error = 0;
 696                 }
 697         } else {
 698                 /*
 699                  * If the record is on-disk we have to queue the deletion by
 700                  * the record's key.  This also causes lookups to skip the
 701                  * record.
 702                  */
 703                 KKASSERT(dip->flags &
 704                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 705                 record = hammer_alloc_mem_record(dip, 0);
 706                 record->type = HAMMER_MEM_RECORD_DEL;
 707                 record->leaf.base = cursor->leaf->base;
 708
 709                 record->target_ip = ip;
 710                 record->flush_state = HAMMER_FST_SETUP;
 711                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 712
 713                 /*
 714                  * The inode now has a dependancy and must be taken out of
 715                  * the idle state.  An inode not in an idle state is given
 716                  * an extra reference.
 717                  */
 718                 if (ip->flush_state == HAMMER_FST_IDLE) {
 719                         hammer_ref(&ip->lock);
 720                         ip->flush_state = HAMMER_FST_SETUP;
 721                 }
 722
 723                 error = hammer_mem_add(record);
 724         }
 725
 726         /*
 727          * One less link.  The file may still be open in the OS even after
 728          * all links have gone away.
 729          *
 730          * We have to terminate the cursor before syncing the inode to
 731          * avoid deadlocking against ourselves.  XXX this may no longer
 732          * be true.
 733          *
 734          * If nlinks drops to zero and the vnode is inactive (or there is
 735          * no vnode), call hammer_inode_unloadable_check() to zonk the
 736          * inode.  If we don't do this here the inode will not be destroyed
 737          * on-media until we unmount.
 738          */
 739         if (error == 0) {
 740                 --ip->ino_data.nlinks;
 741                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 742                 if (ip->ino_data.nlinks == 0 &&
 743                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 744                         hammer_done_cursor(cursor);
 745                         hammer_inode_unloadable_check(ip, 1);
 746                         hammer_flush_inode(ip, 0);
 747                 }
 748
 749         }
 750         return(error);
 751 }
 752
 753 /*
 754  * Add a record to an inode.
 755  *
 756  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 757  * initialize the following additional fields:
 758  *
 759  * The related inode should be share-locked by the caller.  The caller is
 760  * on the frontend.
 761  *
 762  * record->rec.entry.base.base.key
 763  * record->rec.entry.base.base.rec_type
 764  * record->rec.entry.base.base.data_len
 765  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 766  */
 767 int
 768 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 769 {
 770         hammer_inode_t ip = record->ip;
 771         int error;
 772
 773         KKASSERT(record->leaf.base.localization != 0);
 774         record->leaf.base.obj_id = ip->obj_id;
 775         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 776         error = hammer_mem_add(record);
 777         return(error);
 778 }
 779
 780 /*
 781  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 782  * reserved so the front-end can flush large data writes without having
 783  * to queue the BIO to the flusher.  Only the related record gets queued
 784  * to the flusher.
 785  */
 786 static hammer_record_t
 787 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 788 {
 789         hammer_record_t record;
 790         struct hammer_btree_leaf_elm leaf;
 791
 792         bzero(&leaf, sizeof(leaf));
 793         leaf.base.obj_id = ip->obj_id;
 794         leaf.base.key = file_offset + bytes;
 795         leaf.base.create_tid = 0;
 796         leaf.base.delete_tid = 0;
 797         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 798         leaf.base.obj_type = 0;                 /* unused */
 799         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
 800         leaf.base.localization = ip->obj_localization + HAMMER_LOCALIZE_MISC;
 801         leaf.data_len = bytes;
 802
 803         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
 804         if (record)
 805                 hammer_ref(&record->lock);
 806         return(record);
 807 }
 808
 809 /*
 810  * Reserve blockmap space placemarked with an in-memory record.
 811  *
 812  * This routine is called by the frontend in order to be able to directly
 813  * flush a buffer cache buffer.  The frontend has locked the related buffer
 814  * cache buffers and we should be able to manipulate any overlapping
 815  * in-memory records.
 816  */
 817 hammer_record_t
 818 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 819                    int *errorp)
 820 {
 821         hammer_record_t record;
 822         hammer_record_t conflict;
 823         int zone;
 824         int flags;
 825
 826         /*
 827          * Deal with conflicting in-memory records.  We cannot have multiple
 828          * in-memory records for the same offset without seriously confusing
 829          * the backend, including but not limited to the backend issuing
 830          * delete-create-delete sequences and asserting on the delete_tid
 831          * being the same as the create_tid.
 832          *
 833          * If we encounter a record with the backend interlock set we cannot
 834          * immediately delete it without confusing the backend.
 835          */
 836         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 837                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 838                         conflict->flags |= HAMMER_RECF_WANTED;
 839                         tsleep(conflict, 0, "hmrrc3", 0);
 840                 } else {
 841                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 842                 }
 843                 hammer_rel_mem_record(conflict);
 844         }
 845
 846         /*
 847          * Create a record to cover the direct write.  This is called with
 848          * the related BIO locked so there should be no possible conflict.
 849          *
 850          * The backend is responsible for finalizing the space reserved in
 851          * this record.
 852          *
 853          * XXX bytes not aligned, depend on the reservation code to
 854          * align the reservation.
 855          */
 856         record = hammer_alloc_mem_record(ip, 0);
 857         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 858                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 859         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 860                                                &record->leaf.data_offset,
 861                                                errorp);
 862         if (record->resv == NULL) {
 863                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 864                 hammer_rel_mem_record(record);
 865                 return(NULL);
 866         }
 867         record->type = HAMMER_MEM_RECORD_DATA;
 868         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 869         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 870         record->leaf.base.obj_id = ip->obj_id;
 871         record->leaf.base.key = file_offset + bytes;
 872         record->leaf.base.localization = ip->obj_localization +
 873                                          HAMMER_LOCALIZE_MISC;
 874         record->leaf.data_len = bytes;
 875         hammer_crc_set_leaf(data, &record->leaf);
 876         flags = record->flags;
 877
 878         hammer_ref(&record->lock);      /* mem_add eats a reference */
 879         *errorp = hammer_mem_add(record);
 880         if (*errorp) {
 881                 conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 882                 kprintf("hammer_ip_add_bulk: error %d conflict %p file_offset %lld bytes %d\n",
 883                         *errorp, conflict, file_offset, bytes);
 884                 if (conflict)
 885                         kprintf("conflict %lld %d\n", conflict->leaf.base.key, conflict->leaf.data_len);
 886                 if (conflict)
 887                         hammer_rel_mem_record(conflict);
 888         }
 889         KKASSERT(*errorp == 0);
 890         conflict = hammer_ip_get_bulk(ip, file_offset, bytes);
 891         if (conflict != record) {
 892                 kprintf("conflict mismatch %p %p %08x\n", conflict, record, record->flags);
 893                 if (conflict)
 894                     kprintf("conflict mismatch %lld/%d %lld/%d\n", conflict->leaf.base.key, conflict->leaf.data_len, record->leaf.base.key, record->leaf.data_len);
 895         }
 896         KKASSERT(conflict == record);
 897         hammer_rel_mem_record(conflict);
 898
 899         return (record);
 900 }
 901
 902 /*
 903  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 904  * and records in a flushing state are handled by the backend.  The vnops
 905  * setattr code will handle the block containing the truncation point.
 906  *
 907  * Partial blocks are not deleted.
 908  */
 909 int
 910 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 911 {
 912         struct rec_trunc_info info;
 913
 914         switch(ip->ino_data.obj_type) {
 915         case HAMMER_OBJTYPE_REGFILE:
 916                 info.rec_type = HAMMER_RECTYPE_DATA;
 917                 break;
 918         case HAMMER_OBJTYPE_DBFILE:
 919                 info.rec_type = HAMMER_RECTYPE_DB;
 920                 break;
 921         default:
 922                 return(EINVAL);
 923         }
 924         info.trunc_off = file_size;
 925         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 926                                    hammer_frontend_trunc_callback, &info);
 927         return(0);
 928 }
 929
 930 static int
 931 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
 932 {
 933         if (record->flags & HAMMER_RECF_DELETED_FE)
 934                 return(0);
 935         if (record->flush_state == HAMMER_FST_FLUSH)
 936                 return(0);
 937         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 938         hammer_ref(&record->lock);
 939         record->flags |= HAMMER_RECF_DELETED_FE;
 940         hammer_rel_mem_record(record);
 941         return(0);
 942 }
 943
 944 /*
 945  * Return 1 if the caller must check for and delete existing records
 946  * before writing out a new data record.
 947  *
 948  * Return 0 if the caller can just insert the record into the B-Tree without
 949  * checking.
 950  */
 951 static int
 952 hammer_record_needs_overwrite_delete(hammer_record_t record)
 953 {
 954         hammer_inode_t ip = record->ip;
 955         int64_t file_offset;
 956         int r;
 957
 958         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
 959                 file_offset = record->leaf.base.key;
 960         else
 961                 file_offset = record->leaf.base.key - record->leaf.data_len;
 962         r = (file_offset < ip->save_trunc_off);
 963         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
 964                 if (ip->save_trunc_off <= record->leaf.base.key)
 965                         ip->save_trunc_off = record->leaf.base.key + 1;
 966         } else {
 967                 if (ip->save_trunc_off < record->leaf.base.key)
 968                         ip->save_trunc_off = record->leaf.base.key;
 969         }
 970         return(r);
 971 }
 972
 973 /*
 974  * Backend code.  Sync a record to the media.
 975  */
 976 int
 977 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
 978 {
 979         hammer_transaction_t trans = cursor->trans;
 980         int64_t file_offset;
 981         int bytes;
 982         void *bdata;
 983         int error;
 984         int doprop;
 985
 986         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 987         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 988         KKASSERT(record->leaf.base.localization != 0);
 989
 990         /*
 991          * If this is a bulk-data record placemarker there may be an existing
 992          * record on-disk, indicating a data overwrite.  If there is the
 993          * on-disk record must be deleted before we can insert our new record.
 994          *
 995          * We've synthesized this record and do not know what the create_tid
 996          * on-disk is, nor how much data it represents.
 997          *
 998          * Keep in mind that (key) for data records is (base_offset + len),
 999          * not (base_offset).  Also, we only want to get rid of on-disk
1000          * records since we are trying to sync our in-memory record, call
1001          * hammer_ip_delete_range() with truncating set to 1 to make sure
1002          * it skips in-memory records.
1003          *
1004          * It is ok for the lookup to return ENOENT.
1005          *
1006          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1007          * to call hammer_ip_delete_range() or not.  This also means we must
1008          * update sync_trunc_off() as we write.
1009          */
1010         if (record->type == HAMMER_MEM_RECORD_DATA &&
1011             hammer_record_needs_overwrite_delete(record)) {
1012                 file_offset = record->leaf.base.key - record->leaf.data_len;
1013                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1014                         ~HAMMER_BUFMASK;
1015                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1016                 error = hammer_ip_delete_range(
1017                                 cursor, record->ip,
1018                                 file_offset, file_offset + bytes - 1,
1019                                 1);
1020                 if (error && error != ENOENT)
1021                         goto done;
1022         }
1023
1024         /*
1025          * If this is a general record there may be an on-disk version
1026          * that must be deleted before we can insert the new record.
1027          */
1028         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1029                 error = hammer_delete_general(cursor, record->ip,
1030                                               &record->leaf);
1031                 if (error && error != ENOENT)
1032                         goto done;
1033         }
1034
1035         /*
1036          * Setup the cursor.
1037          */
1038         hammer_normalize_cursor(cursor);
1039         cursor->key_beg = record->leaf.base;
1040         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1041         cursor->flags |= HAMMER_CURSOR_BACKEND;
1042         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1043
1044         /*
1045          * Records can wind up on-media before the inode itself is on-media.
1046          * Flag the case.
1047          */
1048         record->ip->flags |= HAMMER_INODE_DONDISK;
1049
1050         /*
1051          * If we are deleting a directory entry an exact match must be
1052          * found on-disk.
1053          */
1054         if (record->type == HAMMER_MEM_RECORD_DEL) {
1055                 error = hammer_btree_lookup(cursor);
1056                 if (error == 0) {
1057                         /* XXX iprec? */
1058                         error = hammer_ip_delete_record(cursor, record->ip,
1059                                                         trans->tid);
1060                         if (error == 0) {
1061                                 record->flags |= HAMMER_RECF_DELETED_FE;
1062                                 record->flags |= HAMMER_RECF_DELETED_BE;
1063                         }
1064                 }
1065                 goto done;
1066         }
1067
1068         /*
1069          * We are inserting.
1070          *
1071          * Issue a lookup to position the cursor and locate the cluster.  The
1072          * target key should not exist.  If we are creating a directory entry
1073          * we may have to iterate the low 32 bits of the key to find an unused
1074          * key.
1075          */
1076         hammer_sync_lock_sh(trans);
1077         cursor->flags |= HAMMER_CURSOR_INSERT;
1078         error = hammer_btree_lookup(cursor);
1079         if (hammer_debug_inode)
1080                 kprintf("DOINSERT LOOKUP %d\n", error);
1081         if (error == 0) {
1082                 kprintf("hammer_ip_sync_record: duplicate rec "
1083                         "at (%016llx)\n", record->leaf.base.key);
1084                 Debugger("duplicate record1");
1085                 error = EIO;
1086         }
1087 #if 0
1088         if (record->type == HAMMER_MEM_RECORD_DATA)
1089                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1090                         record->leaf.base.key - record->leaf.data_len,
1091                         record->leaf.data_offset, error);
1092 #endif
1093
1094         if (error != ENOENT)
1095                 goto done_unlock;
1096
1097         /*
1098          * Allocate the record and data.  The result buffers will be
1099          * marked as being modified and further calls to
1100          * hammer_modify_buffer() will result in unneeded UNDO records.
1101          *
1102          * Support zero-fill records (data == NULL and data_len != 0)
1103          */
1104         if (record->type == HAMMER_MEM_RECORD_DATA) {
1105                 /*
1106                  * The data portion of a bulk-data record has already been
1107                  * committed to disk, we need only adjust the layer2
1108                  * statistics in the same transaction as our B-Tree insert.
1109                  */
1110                 KKASSERT(record->leaf.data_offset != 0);
1111                 hammer_blockmap_finalize(trans, record->leaf.data_offset,
1112                                          record->leaf.data_len);
1113                 error = 0;
1114         } else if (record->data && record->leaf.data_len) {
1115                 /*
1116                  * Wholely cached record, with data.  Allocate the data.
1117                  */
1118                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1119                                           record->leaf.base.rec_type,
1120                                           &record->leaf.data_offset,
1121                                           &cursor->data_buffer, &error);
1122                 if (bdata == NULL)
1123                         goto done_unlock;
1124                 hammer_crc_set_leaf(record->data, &record->leaf);
1125                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1126                 bcopy(record->data, bdata, record->leaf.data_len);
1127                 hammer_modify_buffer_done(cursor->data_buffer);
1128         } else {
1129                 /*
1130                  * Wholely cached record, without data.
1131                  */
1132                 record->leaf.data_offset = 0;
1133                 record->leaf.data_crc = 0;
1134         }
1135
1136         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1137         if (hammer_debug_inode && error)
1138                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1139
1140         /*
1141          * Our record is on-disk, normally mark the in-memory version as
1142          * deleted.  If the record represented a directory deletion but
1143          * we had to sync a valid directory entry to disk we must convert
1144          * the record to a covering delete so the frontend does not have
1145          * visibility on the synced entry.
1146          */
1147         if (error == 0) {
1148                 if (doprop) {
1149                         hammer_btree_do_propagation(cursor,
1150                                                     record->ip->pfsm,
1151                                                     &record->leaf);
1152                 }
1153                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1154                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1155                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1156                         record->type = HAMMER_MEM_RECORD_DEL;
1157                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1158                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1159                         /* hammer_flush_record_done takes care of the rest */
1160                 } else {
1161                         record->flags |= HAMMER_RECF_DELETED_FE;
1162                         record->flags |= HAMMER_RECF_DELETED_BE;
1163                 }
1164         } else {
1165                 if (record->leaf.data_offset) {
1166                         hammer_blockmap_free(trans, record->leaf.data_offset,
1167                                              record->leaf.data_len);
1168                 }
1169         }
1170 done_unlock:
1171         hammer_sync_unlock(trans);
1172 done:
1173         return(error);
1174 }
1175
1176 /*
1177  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1178  * entry's key is used to deal with hash collisions in the upper 32 bits.
1179  * A unique 64 bit key is generated in-memory and may be regenerated a
1180  * second time when the directory record is flushed to the on-disk B-Tree.
1181  *
1182  * A referenced record is passed to this function.  This function
1183  * eats the reference.  If an error occurs the record will be deleted.
1184  *
1185  * A copy of the temporary record->data pointer provided by the caller
1186  * will be made.
1187  */
1188 static
1189 int
1190 hammer_mem_add(hammer_record_t record)
1191 {
1192         hammer_mount_t hmp = record->ip->hmp;
1193
1194         /*
1195          * Make a private copy of record->data
1196          */
1197         if (record->data)
1198                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1199
1200         /*
1201          * Insert into the RB tree.  A unique key should have already
1202          * been selected if this is a directory entry.
1203          */
1204         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1205                 record->flags |= HAMMER_RECF_DELETED_FE;
1206                 hammer_rel_mem_record(record);
1207                 return (EEXIST);
1208         }
1209         ++hmp->count_newrecords;
1210         ++hmp->rsv_recs;
1211         ++record->ip->rsv_recs;
1212         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1213         record->flags |= HAMMER_RECF_ONRBTREE;
1214         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1215         hammer_rel_mem_record(record);
1216         return(0);
1217 }
1218
1219 /************************************************************************
1220  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1221  ************************************************************************
1222  *
1223  * These functions augment the B-Tree scanning functions in hammer_btree.c
1224  * by merging in-memory records with on-disk records.
1225  */
1226
1227 /*
1228  * Locate a particular record either in-memory or on-disk.
1229  *
1230  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1231  * NOT be called to iterate results.
1232  */
1233 int
1234 hammer_ip_lookup(hammer_cursor_t cursor)
1235 {
1236         int error;
1237
1238         /*
1239          * If the element is in-memory return it without searching the
1240          * on-disk B-Tree
1241          */
1242         KKASSERT(cursor->ip);
1243         error = hammer_mem_lookup(cursor);
1244         if (error == 0) {
1245                 cursor->leaf = &cursor->iprec->leaf;
1246                 return(error);
1247         }
1248         if (error != ENOENT)
1249                 return(error);
1250
1251         /*
1252          * If the inode has on-disk components search the on-disk B-Tree.
1253          */
1254         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1255                 return(error);
1256         error = hammer_btree_lookup(cursor);
1257         if (error == 0)
1258                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1259         return(error);
1260 }
1261
1262 /*
1263  * Locate the first record within the cursor's key_beg/key_end range,
1264  * restricted to a particular inode.  0 is returned on success, ENOENT
1265  * if no records matched the requested range, or some other error.
1266  *
1267  * When 0 is returned hammer_ip_next() may be used to iterate additional
1268  * records within the requested range.
1269  *
1270  * This function can return EDEADLK, requiring the caller to terminate
1271  * the cursor and try again.
1272  */
1273 int
1274 hammer_ip_first(hammer_cursor_t cursor)
1275 {
1276         hammer_inode_t ip = cursor->ip;
1277         int error;
1278
1279         KKASSERT(ip != NULL);
1280
1281         /*
1282          * Clean up fields and setup for merged scan
1283          */
1284         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1285         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1286         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1287         if (cursor->iprec) {
1288                 hammer_rel_mem_record(cursor->iprec);
1289                 cursor->iprec = NULL;
1290         }
1291
1292         /*
1293          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1294          * exact lookup so if we get ENOENT we have to call the iterate
1295          * function to validate the first record after the begin key.
1296          *
1297          * The ATEDISK flag is used by hammer_btree_iterate to determine
1298          * whether it must index forwards or not.  It is also used here
1299          * to select the next record from in-memory or on-disk.
1300          *
1301          * EDEADLK can only occur if the lookup hit an empty internal
1302          * element and couldn't delete it.  Since this could only occur
1303          * in-range, we can just iterate from the failure point.
1304          */
1305         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1306                 error = hammer_btree_lookup(cursor);
1307                 if (error == ENOENT || error == EDEADLK) {
1308                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1309                         if (hammer_debug_general & 0x2000)
1310                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1311                         error = hammer_btree_iterate(cursor);
1312                 }
1313                 if (error && error != ENOENT)
1314                         return(error);
1315                 if (error == 0) {
1316                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1317                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1318                 } else {
1319                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1320                 }
1321         }
1322
1323         /*
1324          * Search the in-memory record list (Red-Black tree).  Unlike the
1325          * B-Tree search, mem_first checks for records in the range.
1326          */
1327         error = hammer_mem_first(cursor);
1328         if (error && error != ENOENT)
1329                 return(error);
1330         if (error == 0) {
1331                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1332                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1333                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1334                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1335         }
1336
1337         /*
1338          * This will return the first matching record.
1339          */
1340         return(hammer_ip_next(cursor));
1341 }
1342
1343 /*
1344  * Retrieve the next record in a merged iteration within the bounds of the
1345  * cursor.  This call may be made multiple times after the cursor has been
1346  * initially searched with hammer_ip_first().
1347  *
1348  * 0 is returned on success, ENOENT if no further records match the
1349  * requested range, or some other error code is returned.
1350  */
1351 int
1352 hammer_ip_next(hammer_cursor_t cursor)
1353 {
1354         hammer_btree_elm_t elm;
1355         hammer_record_t rec, save;
1356         int error;
1357         int r;
1358
1359 next_btree:
1360         /*
1361          * Load the current on-disk and in-memory record.  If we ate any
1362          * records we have to get the next one.
1363          *
1364          * If we deleted the last on-disk record we had scanned ATEDISK will
1365          * be clear and DELBTREE will be set, forcing a call to iterate. The
1366          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1367          * element.  If ATEDISK is set, iterate will skip the 'current'
1368          * element.
1369          *
1370          * Get the next on-disk record
1371          */
1372         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1373                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1374                         error = hammer_btree_iterate(cursor);
1375                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1376                         if (error == 0) {
1377                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1378                                 hammer_cache_node(&cursor->ip->cache[1],
1379                                                   cursor->node);
1380                         } else {
1381                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1382                                                  HAMMER_CURSOR_ATEDISK;
1383                         }
1384                 }
1385         }
1386
1387 next_memory:
1388         /*
1389          * Get the next in-memory record.
1390          *
1391          * hammer_rec_scan_cmp:  Is the record still in our general range,
1392          *                       (non-inclusive of snapshot exclusions)?
1393          * hammer_rec_scan_callback: Is the record in our snapshot?
1394          */
1395         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1396                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1397                         save = cursor->iprec;
1398                         cursor->iprec = NULL;
1399                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1400                         while (rec) {
1401                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1402                                         break;
1403                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1404                                         break;
1405                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1406                         }
1407                         if (save)
1408                                 hammer_rel_mem_record(save);
1409                         if (cursor->iprec) {
1410                                 KKASSERT(cursor->iprec == rec);
1411                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1412                         } else {
1413                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1414                         }
1415                 }
1416         }
1417
1418         /*
1419          * The memory record may have become stale while being held in
1420          * cursor->iprec.  We are interlocked against the backend on
1421          * with regards to B-Tree entries.
1422          */
1423         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1424                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1425                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1426                         goto next_memory;
1427                 }
1428         }
1429
1430         /*
1431          * Extract either the disk or memory record depending on their
1432          * relative position.
1433          */
1434         error = 0;
1435         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1436         case 0:
1437                 /*
1438                  * Both entries valid.   Compare the entries and nominally
1439                  * return the first one in the sort order.  Numerous cases
1440                  * require special attention, however.
1441                  */
1442                 elm = &cursor->node->ondisk->elms[cursor->index];
1443                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1444
1445                 /*
1446                  * If the two entries differ only by their key (-2/2) or
1447                  * create_tid (-1/1), and are DATA records, we may have a
1448                  * nominal match.  We have to calculate the base file
1449                  * offset of the data.
1450                  */
1451                 if (r <= 2 && r >= -2 && r != 0 &&
1452                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1453                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1454                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1455                         int64_t base2 = cursor->iprec->leaf.base.key -
1456                                         cursor->iprec->leaf.data_len;
1457                         if (base1 == base2)
1458                                 r = 0;
1459                 }
1460
1461                 if (r < 0) {
1462                         error = hammer_btree_extract(cursor,
1463                                                      HAMMER_CURSOR_GET_LEAF);
1464                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1465                         break;
1466                 }
1467
1468                 /*
1469                  * If the entries match exactly the memory entry is either
1470                  * an on-disk directory entry deletion or a bulk data
1471                  * overwrite.  If it is a directory entry deletion we eat
1472                  * both entries.
1473                  *
1474                  * For the bulk-data overwrite case it is possible to have
1475                  * visibility into both, which simply means the syncer
1476                  * hasn't gotten around to doing the delete+insert sequence
1477                  * on the B-Tree.  Use the memory entry and throw away the
1478                  * on-disk entry.
1479                  *
1480                  * If the in-memory record is not either of these we
1481                  * probably caught the syncer while it was syncing it to
1482                  * the media.  Since we hold a shared lock on the cursor,
1483                  * the in-memory record had better be marked deleted at
1484                  * this point.
1485                  */
1486                 if (r == 0) {
1487                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1488                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1489                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1490                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1491                                         goto next_btree;
1492                                 }
1493                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1494                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1495                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1496                                 }
1497                                 /* fall through to memory entry */
1498                         } else {
1499                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1500                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1501                                 goto next_memory;
1502                         }
1503                 }
1504                 /* fall through to the memory entry */
1505         case HAMMER_CURSOR_ATEDISK:
1506                 /*
1507                  * Only the memory entry is valid.
1508                  */
1509                 cursor->leaf = &cursor->iprec->leaf;
1510                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1511
1512                 /*
1513                  * If the memory entry is an on-disk deletion we should have
1514                  * also had found a B-Tree record.  If the backend beat us
1515                  * to it it would have interlocked the cursor and we should
1516                  * have seen the in-memory record marked DELETED_FE.
1517                  */
1518                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1519                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1520                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1521                 }
1522                 break;
1523         case HAMMER_CURSOR_ATEMEM:
1524                 /*
1525                  * Only the disk entry is valid
1526                  */
1527                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1528                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1529                 break;
1530         default:
1531                 /*
1532                  * Neither entry is valid
1533                  *
1534                  * XXX error not set properly
1535                  */
1536                 cursor->leaf = NULL;
1537                 error = ENOENT;
1538                 break;
1539         }
1540         return(error);
1541 }
1542
1543 /*
1544  * Resolve the cursor->data pointer for the current cursor position in
1545  * a merged iteration.
1546  */
1547 int
1548 hammer_ip_resolve_data(hammer_cursor_t cursor)
1549 {
1550         hammer_record_t record;
1551         int error;
1552
1553         if (hammer_cursor_inmem(cursor)) {
1554                 /*
1555                  * The data associated with an in-memory record is usually
1556                  * kmalloced, but reserve-ahead data records will have an
1557                  * on-disk reference.
1558                  *
1559                  * NOTE: Reserve-ahead data records must be handled in the
1560                  * context of the related high level buffer cache buffer
1561                  * to interlock against async writes.
1562                  */
1563                 record = cursor->iprec;
1564                 cursor->data = record->data;
1565                 error = 0;
1566                 if (cursor->data == NULL) {
1567                         KKASSERT(record->leaf.base.rec_type ==
1568                                  HAMMER_RECTYPE_DATA);
1569                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1570                                                     record->leaf.data_offset,
1571                                                     record->leaf.data_len,
1572                                                     &error,
1573                                                     &cursor->data_buffer);
1574                 }
1575         } else {
1576                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1577                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1578         }
1579         return(error);
1580 }
1581
1582 /*
1583  * Backend truncation / record replacement - delete records in range.
1584  *
1585  * Delete all records within the specified range for inode ip.  In-memory
1586  * records still associated with the frontend are ignored.
1587  *
1588  * If truncating is non-zero in-memory records associated with the back-end
1589  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1590  *
1591  * NOTES:
1592  *
1593  *      * An unaligned range will cause new records to be added to cover
1594  *        the edge cases. (XXX not implemented yet).
1595  *
1596  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1597  *        also do not deal with unaligned ranges.
1598  *
1599  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1600  *
1601  *      * Record keys for regular file data have to be special-cased since
1602  *        they indicate the end of the range (key = base + bytes).
1603  *
1604  *      * This function may be asked to delete ridiculously huge ranges, for
1605  *        example if someone truncates or removes a 1TB regular file.  We
1606  *        must be very careful on restarts and we may have to stop w/
1607  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1608  */
1609 int
1610 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1611                        int64_t ran_beg, int64_t ran_end, int truncating)
1612 {
1613         hammer_transaction_t trans = cursor->trans;
1614         hammer_btree_leaf_elm_t leaf;
1615         int error;
1616         int64_t off;
1617         int64_t tmp64;
1618
1619 #if 0
1620         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1621 #endif
1622
1623         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1624 retry:
1625         hammer_normalize_cursor(cursor);
1626         cursor->key_beg.localization = ip->obj_localization +
1627                                        HAMMER_LOCALIZE_MISC;
1628         cursor->key_beg.obj_id = ip->obj_id;
1629         cursor->key_beg.create_tid = 0;
1630         cursor->key_beg.delete_tid = 0;
1631         cursor->key_beg.obj_type = 0;
1632
1633         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1634                 cursor->key_beg.key = ran_beg;
1635                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1636         } else {
1637                 /*
1638                  * The key in the B-Tree is (base+bytes), so the first possible
1639                  * matching key is ran_beg + 1.
1640                  */
1641                 cursor->key_beg.key = ran_beg + 1;
1642                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1643         }
1644
1645         cursor->key_end = cursor->key_beg;
1646         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1647                 cursor->key_end.key = ran_end;
1648         } else {
1649                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1650                 if (tmp64 < ran_end)
1651                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1652                 else
1653                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1654         }
1655
1656         cursor->asof = ip->obj_asof;
1657         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1658         cursor->flags |= HAMMER_CURSOR_ASOF;
1659         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1660         cursor->flags |= HAMMER_CURSOR_BACKEND;
1661         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1662
1663         error = hammer_ip_first(cursor);
1664
1665         /*
1666          * Iterate through matching records and mark them as deleted.
1667          */
1668         while (error == 0) {
1669                 leaf = cursor->leaf;
1670
1671                 KKASSERT(leaf->base.delete_tid == 0);
1672                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1673
1674                 /*
1675                  * There may be overlap cases for regular file data.  Also
1676                  * remember the key for a regular file record is (base + len),
1677                  * NOT (base).
1678                  *
1679                  * Note that do to duplicates (mem & media) allowed by
1680                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1681                  */
1682                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1683                         off = leaf->base.key - leaf->data_len;
1684                         /*
1685                          * Check the left edge case.  We currently do not
1686                          * split existing records.
1687                          */
1688                         if (off < ran_beg && leaf->base.key > ran_beg) {
1689                                 panic("hammer left edge case %016llx %d\n",
1690                                         leaf->base.key, leaf->data_len);
1691                         }
1692
1693                         /*
1694                          * Check the right edge case.  Note that the
1695                          * record can be completely out of bounds, which
1696                          * terminates the search.
1697                          *
1698                          * base->key is exclusive of the right edge while
1699                          * ran_end is inclusive of the right edge.  The
1700                          * (key - data_len) left boundary is inclusive.
1701                          *
1702                          * XXX theory-check this test at some point, are
1703                          * we missing a + 1 somewhere?  Note that ran_end
1704                          * could overflow.
1705                          */
1706                         if (leaf->base.key - 1 > ran_end) {
1707                                 if (leaf->base.key - leaf->data_len > ran_end)
1708                                         break;
1709                                 panic("hammer right edge case\n");
1710                         }
1711                 } else {
1712                         off = leaf->base.key;
1713                 }
1714
1715                 /*
1716                  * Delete the record.  When truncating we do not delete
1717                  * in-memory (data) records because they represent data
1718                  * written after the truncation.
1719                  *
1720                  * This will also physically destroy the B-Tree entry and
1721                  * data if the retention policy dictates.  The function
1722                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1723                  * uses to perform a fixup.
1724                  */
1725                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1726                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1727                         /*
1728                          * If we have built up too many meta-buffers we risk
1729                          * deadlocking the kernel and must stop.  This can
1730                          * occur when deleting ridiculously huge files.
1731                          * sync_trunc_off is updated so the next cycle does
1732                          * not re-iterate records we have already deleted.
1733                          *
1734                          * This is only done with formal truncations.
1735                          */
1736                         if (truncating > 1 && error == 0 &&
1737                             hammer_flusher_meta_limit(ip->hmp)) {
1738                                 ip->sync_trunc_off = off;
1739                                 error = EWOULDBLOCK;
1740                         }
1741                 }
1742                 if (error)
1743                         break;
1744                 ran_beg = off;  /* for restart */
1745                 error = hammer_ip_next(cursor);
1746         }
1747         if (cursor->node)
1748                 hammer_cache_node(&ip->cache[1], cursor->node);
1749
1750         if (error == EDEADLK) {
1751                 hammer_done_cursor(cursor);
1752                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1753                 if (error == 0)
1754                         goto retry;
1755         }
1756         if (error == ENOENT)
1757                 error = 0;
1758         return(error);
1759 }
1760
1761 /*
1762  * This backend function deletes the specified record on-disk, similar to
1763  * delete_range but for a specific record.  Unlike the exact deletions
1764  * used when deleting a directory entry this function uses an ASOF search
1765  * like delete_range.
1766  *
1767  * This function may be called with ip->obj_asof set for a slave snapshot,
1768  * so don't use it.  We always delete non-historical records only.
1769  */
1770 static int
1771 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
1772                       hammer_btree_leaf_elm_t leaf)
1773 {
1774         hammer_transaction_t trans = cursor->trans;
1775         int error;
1776
1777         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1778 retry:
1779         hammer_normalize_cursor(cursor);
1780         cursor->key_beg = leaf->base;
1781         cursor->asof = HAMMER_MAX_TID;
1782         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1783         cursor->flags |= HAMMER_CURSOR_ASOF;
1784         cursor->flags |= HAMMER_CURSOR_BACKEND;
1785         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1786
1787         error = hammer_btree_lookup(cursor);
1788         if (error == 0) {
1789                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1790         }
1791         if (error == EDEADLK) {
1792                 hammer_done_cursor(cursor);
1793                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1794                 if (error == 0)
1795                         goto retry;
1796         }
1797         return(error);
1798 }
1799
1800 /*
1801  * This function deletes remaining auxillary records when an inode is
1802  * being deleted.  This function explicitly does not delete the
1803  * inode record, directory entry, data, or db records.  Those must be
1804  * properly disposed of prior to this call.
1805  */
1806 int
1807 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
1808 {
1809         hammer_transaction_t trans = cursor->trans;
1810         hammer_btree_leaf_elm_t leaf;
1811         int error;
1812
1813         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1814 retry:
1815         hammer_normalize_cursor(cursor);
1816         cursor->key_beg.localization = ip->obj_localization +
1817                                        HAMMER_LOCALIZE_MISC;
1818         cursor->key_beg.obj_id = ip->obj_id;
1819         cursor->key_beg.create_tid = 0;
1820         cursor->key_beg.delete_tid = 0;
1821         cursor->key_beg.obj_type = 0;
1822         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
1823         cursor->key_beg.key = HAMMER_MIN_KEY;
1824
1825         cursor->key_end = cursor->key_beg;
1826         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
1827         cursor->key_end.key = HAMMER_MAX_KEY;
1828
1829         cursor->asof = ip->obj_asof;
1830         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1831         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1832         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1833         cursor->flags |= HAMMER_CURSOR_BACKEND;
1834
1835         error = hammer_ip_first(cursor);
1836
1837         /*
1838          * Iterate through matching records and mark them as deleted.
1839          */
1840         while (error == 0) {
1841                 leaf = cursor->leaf;
1842
1843                 KKASSERT(leaf->base.delete_tid == 0);
1844
1845                 /*
1846                  * Mark the record and B-Tree entry as deleted.  This will
1847                  * also physically delete the B-Tree entry, record, and
1848                  * data if the retention policy dictates.  The function
1849                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1850                  * uses to perform a fixup.
1851                  *
1852                  * Directory entries (and delete-on-disk directory entries)
1853                  * must be synced and cannot be deleted.
1854                  */
1855                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
1856                 ++*countp;
1857                 if (error)
1858                         break;
1859                 error = hammer_ip_next(cursor);
1860         }
1861         if (cursor->node)
1862                 hammer_cache_node(&ip->cache[1], cursor->node);
1863         if (error == EDEADLK) {
1864                 hammer_done_cursor(cursor);
1865                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1866                 if (error == 0)
1867                         goto retry;
1868         }
1869         if (error == ENOENT)
1870                 error = 0;
1871         return(error);
1872 }
1873
1874 /*
1875  * Delete the record at the current cursor.  On success the cursor will
1876  * be positioned appropriately for an iteration but may no longer be at
1877  * a leaf node.
1878  *
1879  * This routine is only called from the backend.
1880  *
1881  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1882  * cursor and retry.
1883  */
1884 int
1885 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1886                         hammer_tid_t tid)
1887 {
1888         hammer_off_t zone2_offset;
1889         hammer_record_t iprec;
1890         hammer_btree_elm_t elm;
1891         hammer_mount_t hmp;
1892         int error;
1893
1894         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1895         KKASSERT(tid != 0);
1896         hmp = cursor->node->hmp;
1897
1898         /*
1899          * In-memory (unsynchronized) records can simply be freed.  This
1900          * only occurs in range iterations since all other records are
1901          * individually synchronized.  Thus there should be no confusion with
1902          * the interlock.
1903          *
1904          * An in-memory record may be deleted before being committed to disk,
1905          * but could have been accessed in the mean time.  The backing store
1906          * may never been marked allocated and so hammer_blockmap_free() may
1907          * never get called on it.  Because of this we have to make sure that
1908          * we've gotten rid of any related hammer_buffer or buffer cache
1909          * buffer.
1910          */
1911         if (hammer_cursor_inmem(cursor)) {
1912                 iprec = cursor->iprec;
1913                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1914                 iprec->flags |= HAMMER_RECF_DELETED_FE;
1915                 iprec->flags |= HAMMER_RECF_DELETED_BE;
1916
1917                 if (iprec->leaf.data_offset && iprec->leaf.data_len) {
1918                         zone2_offset = hammer_blockmap_lookup(hmp, iprec->leaf.data_offset, &error);
1919                         KKASSERT(error == 0);
1920                         hammer_del_buffers(hmp,
1921                                            iprec->leaf.data_offset,
1922                                            zone2_offset,
1923                                            iprec->leaf.data_len);
1924                 }
1925                 return(0);
1926         }
1927
1928         /*
1929          * On-disk records are marked as deleted by updating their delete_tid.
1930          * This does not effect their position in the B-Tree (which is based
1931          * on their create_tid).
1932          *
1933          * Frontend B-Tree operations track inodes so we tell
1934          * hammer_delete_at_cursor() not to.
1935          */
1936         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1937         elm = NULL;
1938
1939         if (error == 0) {
1940                 error = hammer_delete_at_cursor(
1941                                 cursor,
1942                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
1943                                 cursor->trans->tid,
1944                                 cursor->trans->time32,
1945                                 0, NULL);
1946         }
1947         return(error);
1948 }
1949
1950 /*
1951  * Delete the B-Tree element at the current cursor and do any necessary
1952  * mirror propagation.
1953  *
1954  * The cursor must be properly positioned for an iteration on return but
1955  * may be pointing at an internal element.
1956  *
1957  * An element can be un-deleted by passing a delete_tid of 0 with
1958  * HAMMER_DELETE_ADJUST.
1959  */
1960 int
1961 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
1962                         hammer_tid_t delete_tid, u_int32_t delete_ts,
1963                         int track, int64_t *stat_bytes)
1964 {
1965         struct hammer_btree_leaf_elm save_leaf;
1966         hammer_transaction_t trans;
1967         hammer_btree_leaf_elm_t leaf;
1968         hammer_node_t node;
1969         hammer_btree_elm_t elm;
1970         hammer_off_t data_offset;
1971         int32_t data_len;
1972         u_int16_t rec_type;
1973         int error;
1974         int icount;
1975         int doprop;
1976
1977         error = hammer_cursor_upgrade(cursor);
1978         if (error)
1979                 return(error);
1980
1981         trans = cursor->trans;
1982         node = cursor->node;
1983         elm = &node->ondisk->elms[cursor->index];
1984         leaf = &elm->leaf;
1985         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1986
1987         hammer_sync_lock_sh(trans);
1988         doprop = 0;
1989         icount = 0;
1990
1991         /*
1992          * Adjust the delete_tid.  Update the mirror_tid propagation field
1993          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
1994          */
1995         if (delete_flags & HAMMER_DELETE_ADJUST) {
1996                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
1997                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
1998                                 icount = -1;
1999                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2000                                 icount = 1;
2001                 }
2002
2003                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2004                 elm->leaf.base.delete_tid = delete_tid;
2005                 elm->leaf.delete_ts = delete_ts;
2006                 hammer_modify_node_done(node);
2007
2008                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2009                         hammer_modify_node_field(trans, node, mirror_tid);
2010                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2011                         hammer_modify_node_done(node);
2012                         doprop = 1;
2013                         if (hammer_debug_general & 0x0002) {
2014                                 kprintf("delete_at_cursor: propagate %016llx"
2015                                         " @%016llx\n",
2016                                         elm->leaf.base.delete_tid,
2017                                         node->node_offset);
2018                         }
2019                 }
2020
2021                 /*
2022                  * Adjust for the iteration.  We have deleted the current
2023                  * element and want to clear ATEDISK so the iteration does
2024                  * not skip the element after, which now becomes the current
2025                  * element.
2026                  */
2027                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2028                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
2029                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2030                 }
2031
2032                 /*
2033                  * An on-disk record cannot have the same delete_tid
2034                  * as its create_tid.  In a chain of record updates
2035                  * this could result in a duplicate record.
2036                  */
2037                 KKASSERT(elm->leaf.base.delete_tid !=
2038                          elm->leaf.base.create_tid);
2039         }
2040
2041         /*
2042          * Destroy the B-Tree element if asked (typically if a nohistory
2043          * file or mount, or when called by the pruning code).
2044          *
2045          * Adjust the ATEDISK flag to properly support iterations.
2046          */
2047         if (delete_flags & HAMMER_DELETE_DESTROY) {
2048                 data_offset = elm->leaf.data_offset;
2049                 data_len = elm->leaf.data_len;
2050                 rec_type = elm->leaf.base.rec_type;
2051                 if (doprop) {
2052                         save_leaf = elm->leaf;
2053                         leaf = &save_leaf;
2054                 }
2055                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2056                     elm->leaf.base.delete_tid == 0) {
2057                         icount = -1;
2058                 }
2059
2060                 error = hammer_btree_delete(cursor);
2061                 if (error == 0) {
2062                         /*
2063                          * This forces a fixup for the iteration because
2064                          * the cursor is now either sitting at the 'next'
2065                          * element or sitting at the end of a leaf.
2066                          */
2067                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2068                                 cursor->flags |= HAMMER_CURSOR_DELBTREE;
2069                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2070                         }
2071                 }
2072                 if (error == 0) {
2073                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2074                         case HAMMER_ZONE_LARGE_DATA:
2075                         case HAMMER_ZONE_SMALL_DATA:
2076                         case HAMMER_ZONE_META:
2077                                 hammer_blockmap_free(trans,
2078                                                      data_offset, data_len);
2079                                 break;
2080                         default:
2081                                 break;
2082                         }
2083                 }
2084         }
2085
2086         /*
2087          * Track inode count and next_tid.  This is used by the mirroring
2088          * and PFS code.  icount can be negative, zero, or positive.
2089          */
2090         if (error == 0 && track) {
2091                 if (icount) {
2092                         hammer_modify_volume_field(trans, trans->rootvol,
2093                                                    vol0_stat_inodes);
2094                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2095                         hammer_modify_volume_done(trans->rootvol);
2096                 }
2097                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2098                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2099                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2100                         hammer_modify_volume_done(trans->rootvol);
2101                 }
2102         }
2103
2104         /*
2105          * mirror_tid propagation occurs if the node's mirror_tid had to be
2106          * updated while adjusting the delete_tid.
2107          *
2108          * This occurs when deleting even in nohistory mode, but does not
2109          * occur when pruning an already-deleted node.
2110          *
2111          * cursor->ip is NULL when called from the pruning, mirroring,
2112          * and pfs code.  If non-NULL propagation will be conditionalized
2113          * on whether the PFS is in no-history mode or not.
2114          */
2115         if (doprop) {
2116                 if (cursor->ip)
2117                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2118                 else
2119                         hammer_btree_do_propagation(cursor, NULL, leaf);
2120         }
2121         hammer_sync_unlock(trans);
2122         return (error);
2123 }
2124
2125 /*
2126  * Determine whether we can remove a directory.  This routine checks whether
2127  * a directory is empty or not and enforces flush connectivity.
2128  *
2129  * Flush connectivity requires that we block if the target directory is
2130  * currently flushing, otherwise it may not end up in the same flush group.
2131  *
2132  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2133  */
2134 int
2135 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2136 {
2137         struct hammer_cursor cursor;
2138         int error;
2139
2140         /*
2141          * Check directory empty
2142          */
2143         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2144
2145         cursor.key_beg.localization = ip->obj_localization +
2146                                       HAMMER_LOCALIZE_MISC;
2147         cursor.key_beg.obj_id = ip->obj_id;
2148         cursor.key_beg.create_tid = 0;
2149         cursor.key_beg.delete_tid = 0;
2150         cursor.key_beg.obj_type = 0;
2151         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2152         cursor.key_beg.key = HAMMER_MIN_KEY;
2153
2154         cursor.key_end = cursor.key_beg;
2155         cursor.key_end.rec_type = 0xFFFF;
2156         cursor.key_end.key = HAMMER_MAX_KEY;
2157
2158         cursor.asof = ip->obj_asof;
2159         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2160
2161         error = hammer_ip_first(&cursor);
2162         if (error == ENOENT)
2163                 error = 0;
2164         else if (error == 0)
2165                 error = ENOTEMPTY;
2166         hammer_done_cursor(&cursor);
2167         return(error);
2168 }
2169