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 1.67 2008/06/13 00:25:33 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_add(hammer_record_t record);
  40 static int hammer_mem_lookup(hammer_cursor_t cursor);
  41 static int hammer_mem_first(hammer_cursor_t cursor);
  42 static int hammer_rec_trunc_callback(hammer_record_t record,
  43                                 void *data __unused);
  44
  45 struct rec_trunc_info {
  46         u_int16_t       rec_type;
  47         int64_t         trunc_off;
  48 };
  49
  50 /*
  51  * Red-black tree support.  Comparison code for insertion.
  52  */
  53 static int
  54 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  55 {
  56         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  57                 return(-1);
  58         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  59                 return(1);
  60
  61         if (rec1->leaf.base.key < rec2->leaf.base.key)
  62                 return(-1);
  63         if (rec1->leaf.base.key > rec2->leaf.base.key)
  64                 return(1);
  65
  66 #if 0
  67         /*
  68          * XXX create_tid is set during sync, memory records are always
  69          * current.  Do not match against create_tid.
  70          */
  71         if (rec1->leaf.base.create_tid == 0) {
  72                 if (rec2->leaf.base.create_tid == 0)
  73                         return(0);
  74                 return(1);
  75         }
  76         if (rec2->leaf.base.create_tid == 0)
  77                 return(-1);
  78
  79         if (rec1->leaf.base.create_tid < rec2->leaf.base.create_tid)
  80                 return(-1);
  81         if (rec1->leaf.base.create_tid > rec2->leaf.base.create_tid)
  82                 return(1);
  83 #endif
  84
  85         /*
  86          * Never match against an item deleted by the front-end.
  87          */
  88         if (rec1->flags & HAMMER_RECF_DELETED_FE)
  89                 return(1);
  90         if (rec2->flags & HAMMER_RECF_DELETED_FE)
  91                 return(-1);
  92
  93         return(0);
  94 }
  95
  96 /*
  97  * Basic record comparison code similar to hammer_btree_cmp().
  98  */
  99 static int
 100 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 101 {
 102         if (elm->rec_type < rec->leaf.base.rec_type)
 103                 return(-3);
 104         if (elm->rec_type > rec->leaf.base.rec_type)
 105                 return(3);
 106
 107         if (elm->key < rec->leaf.base.key)
 108                 return(-2);
 109         if (elm->key > rec->leaf.base.key)
 110                 return(2);
 111
 112 #if 0
 113         /*
 114          * XXX create_tid is set during sync, memory records are always
 115          * current.  Do not match against create_tid.
 116          */
 117         if (elm->create_tid == 0) {
 118                 if (rec->leaf.base.create_tid == 0)
 119                         return(0);
 120                 return(1);
 121         }
 122         if (rec->leaf.base.create_tid == 0)
 123                 return(-1);
 124         if (elm->create_tid < rec->leaf.base.create_tid)
 125                 return(-1);
 126         if (elm->create_tid > rec->leaf.base.create_tid)
 127                 return(1);
 128 #endif
 129         /*
 130          * Never match against an item deleted by the front-end.
 131          */
 132         if (rec->flags & HAMMER_RECF_DELETED_FE)
 133                 return(1);
 134         return(0);
 135 }
 136
 137 /*
 138  * Special LOOKUP_INFO to locate an overlapping record.  This used by
 139  * the reservation code to implement small-block records (whos keys will
 140  * be different depending on data_len, when representing the same base
 141  * offset).
 142  *
 143  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 144  */
 145 static int
 146 hammer_rec_overlap_compare(hammer_btree_leaf_elm_t leaf, hammer_record_t rec)
 147 {
 148         if (leaf->base.rec_type < rec->leaf.base.rec_type)
 149                 return(-3);
 150         if (leaf->base.rec_type > rec->leaf.base.rec_type)
 151                 return(3);
 152
 153         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 154                 /* leaf_end <= rec_beg */
 155                 if (leaf->base.key <= rec->leaf.base.key - rec->leaf.data_len)
 156                         return(-2);
 157                 /* leaf_beg >= rec_end */
 158                 if (leaf->base.key - leaf->data_len >= rec->leaf.base.key)
 159                         return(2);
 160         } else {
 161                 if (leaf->base.key < rec->leaf.base.key)
 162                         return(-2);
 163                 if (leaf->base.key > rec->leaf.base.key)
 164                         return(2);
 165         }
 166
 167 #if 0
 168         if (leaf->base.create_tid == 0) {
 169                 if (rec->leaf.base.create_tid == 0)
 170                         return(0);
 171                 return(1);
 172         }
 173         if (rec->leaf.base.create_tid == 0)
 174                 return(-1);
 175         if (leaf->base.create_tid < rec->leaf.base.create_tid)
 176                 return(-1);
 177         if (leaf->base.create_tid > rec->leaf.base.create_tid)
 178                 return(1);
 179 #endif
 180         /*
 181          * Never match against an item deleted by the front-end.
 182          */
 183         if (rec->flags & HAMMER_RECF_DELETED_FE)
 184                 return(1);
 185         return(0);
 186 }
 187
 188 /*
 189  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 190  * is reversed so the comparison result has to be negated.  key_beg and
 191  * key_end are both range-inclusive.
 192  *
 193  * Localized deletions are not cached in-memory.
 194  */
 195 static
 196 int
 197 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 198 {
 199         hammer_cursor_t cursor = data;
 200         int r;
 201
 202         r = hammer_rec_cmp(&cursor->key_beg, rec);
 203         if (r > 1)
 204                 return(-1);
 205         r = hammer_rec_cmp(&cursor->key_end, rec);
 206         if (r < -1)
 207                 return(1);
 208         return(0);
 209 }
 210
 211 /*
 212  * This compare function is used when simply looking up key_beg.
 213  */
 214 static
 215 int
 216 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 217 {
 218         hammer_cursor_t cursor = data;
 219         int r;
 220
 221         r = hammer_rec_cmp(&cursor->key_beg, rec);
 222         if (r > 1)
 223                 return(-1);
 224         if (r < -1)
 225                 return(1);
 226         return(0);
 227 }
 228
 229 /*
 230  * Locate blocks within the truncation range.  Partial blocks do not count.
 231  */
 232 static
 233 int
 234 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 235 {
 236         struct rec_trunc_info *info = data;
 237
 238         if (rec->leaf.base.rec_type < info->rec_type)
 239                 return(-1);
 240         if (rec->leaf.base.rec_type > info->rec_type)
 241                 return(1);
 242
 243         switch(rec->leaf.base.rec_type) {
 244         case HAMMER_RECTYPE_DB:
 245                 /*
 246                  * DB record key is not beyond the truncation point, retain.
 247                  */
 248                 if (rec->leaf.base.key < info->trunc_off)
 249                         return(-1);
 250                 break;
 251         case HAMMER_RECTYPE_DATA:
 252                 /*
 253                  * DATA record offset start is not beyond the truncation point,
 254                  * retain.
 255                  */
 256                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 257                         return(-1);
 258                 break;
 259         default:
 260                 panic("hammer_rec_trunc_cmp: unexpected record type");
 261         }
 262
 263         /*
 264          * The record start is >= the truncation point, return match,
 265          * the record should be destroyed.
 266          */
 267         return(0);
 268 }
 269
 270 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 271 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 272                     hammer_rec_overlap_compare, hammer_btree_leaf_elm_t);
 273
 274 /*
 275  * Allocate a record for the caller to finish filling in.  The record is
 276  * returned referenced.
 277  */
 278 hammer_record_t
 279 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 280 {
 281         hammer_record_t record;
 282
 283         ++hammer_count_records;
 284         record = kmalloc(sizeof(*record), M_HAMMER, M_WAITOK | M_ZERO);
 285         record->flush_state = HAMMER_FST_IDLE;
 286         record->ip = ip;
 287         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 288         record->leaf.data_len = data_len;
 289         hammer_ref(&record->lock);
 290
 291         if (data_len) {
 292                 record->data = kmalloc(data_len, M_HAMMER, M_WAITOK | M_ZERO);
 293                 record->flags |= HAMMER_RECF_ALLOCDATA;
 294                 ++hammer_count_record_datas;
 295         }
 296
 297         return (record);
 298 }
 299
 300 void
 301 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 302 {
 303         while (record->flush_state == HAMMER_FST_FLUSH) {
 304                 record->flags |= HAMMER_RECF_WANTED;
 305                 tsleep(record, 0, ident, 0);
 306         }
 307 }
 308
 309 /*
 310  * Called from the backend, hammer_inode.c, after a record has been
 311  * flushed to disk.  The record has been exclusively locked by the
 312  * caller and interlocked with BE.
 313  *
 314  * We clean up the state, unlock, and release the record (the record
 315  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 316  */
 317 void
 318 hammer_flush_record_done(hammer_record_t record, int error)
 319 {
 320         hammer_inode_t target_ip;
 321
 322         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 323         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 324
 325         if (error) {
 326                 /*
 327                  * An error occured, the backend was unable to sync the
 328                  * record to its media.  Leave the record intact.
 329                  */
 330                 Debugger("flush_record_done error");
 331         }
 332
 333         if (record->flags & HAMMER_RECF_DELETED_BE) {
 334                 if ((target_ip = record->target_ip) != NULL) {
 335                         TAILQ_REMOVE(&target_ip->target_list, record,
 336                                      target_entry);
 337                         record->target_ip = NULL;
 338                         hammer_test_inode(target_ip);
 339                 }
 340                 record->flush_state = HAMMER_FST_IDLE;
 341         } else {
 342                 if (record->target_ip) {
 343                         record->flush_state = HAMMER_FST_SETUP;
 344                         hammer_test_inode(record->ip);
 345                         hammer_test_inode(record->target_ip);
 346                 } else {
 347                         record->flush_state = HAMMER_FST_IDLE;
 348                 }
 349         }
 350         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 351         if (record->flags & HAMMER_RECF_WANTED) {
 352                 record->flags &= ~HAMMER_RECF_WANTED;
 353                 wakeup(record);
 354         }
 355         hammer_rel_mem_record(record);
 356 }
 357
 358 /*
 359  * Release a memory record.  Records marked for deletion are immediately
 360  * removed from the RB-Tree but otherwise left intact until the last ref
 361  * goes away.
 362  */
 363 void
 364 hammer_rel_mem_record(struct hammer_record *record)
 365 {
 366         hammer_inode_t ip, target_ip;
 367
 368         hammer_unref(&record->lock);
 369
 370         if (record->lock.refs == 0) {
 371                 /*
 372                  * Upon release of the last reference wakeup any waiters.
 373                  * The record structure may get destroyed so callers will
 374                  * loop up and do a relookup.
 375                  */
 376                 ip = record->ip;
 377
 378                 /*
 379                  * Upon release of the last reference a record marked deleted
 380                  * is destroyed.
 381                  */
 382                 if (record->flags & HAMMER_RECF_DELETED_FE) {
 383                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 384
 385                         if ((target_ip = record->target_ip) != NULL) {
 386                                 TAILQ_REMOVE(&target_ip->target_list,
 387                                              record, target_entry);
 388                                 record->target_ip = NULL;
 389                                 hammer_test_inode(target_ip);
 390                         }
 391
 392                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 393                                 RB_REMOVE(hammer_rec_rb_tree,
 394                                           &record->ip->rec_tree,
 395                                           record);
 396                                 KKASSERT(ip->rsv_recs > 0);
 397                                 --ip->hmp->rsv_recs;
 398                                 --ip->rsv_recs;
 399                                 ip->hmp->rsv_databytes -= record->leaf.data_len;
 400                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 401
 402                                 if ((ip->flags & HAMMER_INODE_PARTIALW) &&
 403                                     ip->rsv_recs <= hammer_limit_irecs) {
 404                                         ip->flags &= ~HAMMER_INODE_PARTIALW;
 405                                         wakeup(&ip->flags);
 406                                 }
 407                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 408                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 409                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 410                                         hammer_test_inode(record->ip);
 411                                 }
 412                         }
 413                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 414                                 --hammer_count_record_datas;
 415                                 kfree(record->data, M_HAMMER);
 416                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 417                         }
 418                         if (record->resv) {
 419                                 hammer_blockmap_reserve_complete(ip->hmp,
 420                                                                  record->resv);
 421                                 record->resv = NULL;
 422                         }
 423                         record->data = NULL;
 424                         --hammer_count_records;
 425                         kfree(record, M_HAMMER);
 426                 }
 427         }
 428 }
 429
 430 /*
 431  * Record visibility depends on whether the record is being accessed by
 432  * the backend or the frontend.
 433  *
 434  * Return non-zero if the record is visible, zero if it isn't or if it is
 435  * deleted.
 436  */
 437 static __inline
 438 int
 439 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 440 {
 441         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 442                 if (record->flags & HAMMER_RECF_DELETED_BE)
 443                         return(0);
 444         } else {
 445                 if (record->flags & HAMMER_RECF_DELETED_FE)
 446                         return(0);
 447         }
 448         return(1);
 449 }
 450
 451 /*
 452  * This callback is used as part of the RB_SCAN function for in-memory
 453  * records.  We terminate it (return -1) as soon as we get a match.
 454  *
 455  * This routine is used by frontend code.
 456  *
 457  * The primary compare code does not account for ASOF lookups.  This
 458  * code handles that case as well as a few others.
 459  */
 460 static
 461 int
 462 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 463 {
 464         hammer_cursor_t cursor = data;
 465
 466         /*
 467          * We terminate on success, so this should be NULL on entry.
 468          */
 469         KKASSERT(cursor->iprec == NULL);
 470
 471         /*
 472          * Skip if the record was marked deleted.
 473          */
 474         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 475                 return(0);
 476
 477         /*
 478          * Skip if not visible due to our as-of TID
 479          */
 480         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 481                 if (cursor->asof < rec->leaf.base.create_tid)
 482                         return(0);
 483                 if (rec->leaf.base.delete_tid &&
 484                     cursor->asof >= rec->leaf.base.delete_tid) {
 485                         return(0);
 486                 }
 487         }
 488
 489         /*
 490          * If the record is queued to the flusher we have to block until
 491          * it isn't.  Otherwise we may see duplication between our memory
 492          * cache and the media.
 493          */
 494         hammer_ref(&rec->lock);
 495
 496 #warning "This deadlocks"
 497 #if 0
 498         if (rec->flush_state == HAMMER_FST_FLUSH)
 499                 hammer_wait_mem_record(rec);
 500 #endif
 501
 502         /*
 503          * The record may have been deleted while we were blocked.
 504          */
 505         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 506                 hammer_rel_mem_record(rec);
 507                 return(0);
 508         }
 509
 510         /*
 511          * Set the matching record and stop the scan.
 512          */
 513         cursor->iprec = rec;
 514         return(-1);
 515 }
 516
 517
 518 /*
 519  * Lookup an in-memory record given the key specified in the cursor.  Works
 520  * just like hammer_btree_lookup() but operates on an inode's in-memory
 521  * record list.
 522  *
 523  * The lookup must fail if the record is marked for deferred deletion.
 524  */
 525 static
 526 int
 527 hammer_mem_lookup(hammer_cursor_t cursor)
 528 {
 529         int error;
 530
 531         KKASSERT(cursor->ip);
 532         if (cursor->iprec) {
 533                 hammer_rel_mem_record(cursor->iprec);
 534                 cursor->iprec = NULL;
 535         }
 536         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 537                                    hammer_rec_scan_callback, cursor);
 538
 539         if (cursor->iprec == NULL)
 540                 error = ENOENT;
 541         else
 542                 error = 0;
 543         return(error);
 544 }
 545
 546 /*
 547  * hammer_mem_first() - locate the first in-memory record matching the
 548  * cursor within the bounds of the key range.
 549  */
 550 static
 551 int
 552 hammer_mem_first(hammer_cursor_t cursor)
 553 {
 554         hammer_inode_t ip;
 555
 556         ip = cursor->ip;
 557         KKASSERT(ip != NULL);
 558
 559         if (cursor->iprec) {
 560                 hammer_rel_mem_record(cursor->iprec);
 561                 cursor->iprec = NULL;
 562         }
 563
 564         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 565                                    hammer_rec_scan_callback, cursor);
 566
 567         /*
 568          * Adjust scan.node and keep it linked into the RB-tree so we can
 569          * hold the cursor through third party modifications of the RB-tree.
 570          */
 571         if (cursor->iprec)
 572                 return(0);
 573         return(ENOENT);
 574 }
 575
 576 void
 577 hammer_mem_done(hammer_cursor_t cursor)
 578 {
 579         if (cursor->iprec) {
 580                 hammer_rel_mem_record(cursor->iprec);
 581                 cursor->iprec = NULL;
 582         }
 583 }
 584
 585 /************************************************************************
 586  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 587  ************************************************************************
 588  *
 589  * These functions manipulate in-memory records.  Such records typically
 590  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 591  */
 592
 593 /*
 594  * Add a directory entry (dip,ncp) which references inode (ip).
 595  *
 596  * Note that the low 32 bits of the namekey are set temporarily to create
 597  * a unique in-memory record, and may be modified a second time when the
 598  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 599  * all 0's when synching to disk, which is not handled here.
 600  */
 601 int
 602 hammer_ip_add_directory(struct hammer_transaction *trans,
 603                      struct hammer_inode *dip, struct namecache *ncp,
 604                      struct hammer_inode *ip)
 605 {
 606         hammer_record_t record;
 607         int error;
 608         int bytes;
 609
 610         bytes = ncp->nc_nlen;   /* NOTE: terminating \0 is NOT included */
 611         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 612         if (++trans->hmp->namekey_iterator == 0)
 613                 ++trans->hmp->namekey_iterator;
 614
 615         record->type = HAMMER_MEM_RECORD_ADD;
 616         record->leaf.base.localization = HAMMER_LOCALIZE_MISC;
 617         record->leaf.base.obj_id = dip->obj_id;
 618         record->leaf.base.key = hammer_directory_namekey(ncp->nc_name, bytes);
 619         record->leaf.base.key += trans->hmp->namekey_iterator;
 620         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 621         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 622         record->data->entry.obj_id = ip->obj_id;
 623         bcopy(ncp->nc_name, record->data->entry.name, bytes);
 624
 625         ++ip->ino_data.nlinks;
 626         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 627
 628         /*
 629          * The target inode and the directory entry are bound together.
 630          */
 631         record->target_ip = ip;
 632         record->flush_state = HAMMER_FST_SETUP;
 633         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 634
 635         /*
 636          * The inode now has a dependancy and must be taken out of the idle
 637          * state.  An inode not in an idle state is given an extra reference.
 638          */
 639         if (ip->flush_state == HAMMER_FST_IDLE) {
 640                 hammer_ref(&ip->lock);
 641                 ip->flush_state = HAMMER_FST_SETUP;
 642         }
 643         error = hammer_mem_add(record);
 644         return(error);
 645 }
 646
 647 /*
 648  * Delete the directory entry and update the inode link count.  The
 649  * cursor must be seeked to the directory entry record being deleted.
 650  *
 651  * The related inode should be share-locked by the caller.  The caller is
 652  * on the frontend.
 653  *
 654  * This function can return EDEADLK requiring the caller to terminate
 655  * the cursor, any locks, wait on the returned record, and retry.
 656  */
 657 int
 658 hammer_ip_del_directory(struct hammer_transaction *trans,
 659                      hammer_cursor_t cursor, struct hammer_inode *dip,
 660                      struct hammer_inode *ip)
 661 {
 662         hammer_record_t record;
 663         int error;
 664
 665         if (hammer_cursor_inmem(cursor)) {
 666                 /*
 667                  * In-memory (unsynchronized) records can simply be freed.
 668                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 669                  * by the backend, we must still avoid races against the
 670                  * backend potentially syncing the record to the media.
 671                  *
 672                  * We cannot call hammer_ip_delete_record(), that routine may
 673                  * only be called from the backend.
 674                  */
 675                 record = cursor->iprec;
 676                 if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
 677                         KKASSERT(cursor->deadlk_rec == NULL);
 678                         hammer_ref(&record->lock);
 679                         cursor->deadlk_rec = record;
 680                         error = EDEADLK;
 681                 } else {
 682                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 683                         record->flags |= HAMMER_RECF_DELETED_FE;
 684                         error = 0;
 685                 }
 686         } else {
 687                 /*
 688                  * If the record is on-disk we have to queue the deletion by
 689                  * the record's key.  This also causes lookups to skip the
 690                  * record.
 691                  */
 692                 KKASSERT(dip->flags &
 693                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 694                 record = hammer_alloc_mem_record(dip, 0);
 695                 record->type = HAMMER_MEM_RECORD_DEL;
 696                 record->leaf.base = cursor->leaf->base;
 697
 698                 record->target_ip = ip;
 699                 record->flush_state = HAMMER_FST_SETUP;
 700                 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 701
 702                 /*
 703                  * The inode now has a dependancy and must be taken out of
 704                  * the idle state.  An inode not in an idle state is given
 705                  * an extra reference.
 706                  */
 707                 if (ip->flush_state == HAMMER_FST_IDLE) {
 708                         hammer_ref(&ip->lock);
 709                         ip->flush_state = HAMMER_FST_SETUP;
 710                 }
 711
 712                 error = hammer_mem_add(record);
 713         }
 714
 715         /*
 716          * One less link.  The file may still be open in the OS even after
 717          * all links have gone away.
 718          *
 719          * We have to terminate the cursor before syncing the inode to
 720          * avoid deadlocking against ourselves.  XXX this may no longer
 721          * be true.
 722          *
 723          * If nlinks drops to zero and the vnode is inactive (or there is
 724          * no vnode), call hammer_inode_unloadable_check() to zonk the
 725          * inode.  If we don't do this here the inode will not be destroyed
 726          * on-media until we unmount.
 727          */
 728         if (error == 0) {
 729                 --ip->ino_data.nlinks;
 730                 hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 731                 if (ip->ino_data.nlinks == 0 &&
 732                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 733                         hammer_done_cursor(cursor);
 734                         hammer_inode_unloadable_check(ip, 1);
 735                         hammer_flush_inode(ip, 0);
 736                 }
 737
 738         }
 739         return(error);
 740 }
 741
 742 /*
 743  * Add a record to an inode.
 744  *
 745  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 746  * initialize the following additional fields:
 747  *
 748  * The related inode should be share-locked by the caller.  The caller is
 749  * on the frontend.
 750  *
 751  * record->rec.entry.base.base.key
 752  * record->rec.entry.base.base.rec_type
 753  * record->rec.entry.base.base.data_len
 754  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 755  */
 756 int
 757 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 758 {
 759         hammer_inode_t ip = record->ip;
 760         int error;
 761
 762         KKASSERT(record->leaf.base.localization != 0);
 763         record->leaf.base.obj_id = ip->obj_id;
 764         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 765         error = hammer_mem_add(record);
 766         return(error);
 767 }
 768
 769 /*
 770  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 771  * reserved so the front-end can flush large data writes without having
 772  * to queue the BIO to the flusher.  Only the related record gets queued
 773  * to the flusher.
 774  */
 775 static hammer_record_t
 776 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 777 {
 778         hammer_record_t record;
 779         struct hammer_btree_leaf_elm leaf;
 780
 781         bzero(&leaf, sizeof(leaf));
 782         leaf.base.obj_id = ip->obj_id;
 783         leaf.base.key = file_offset + bytes;
 784         leaf.base.create_tid = 0;
 785         leaf.base.delete_tid = 0;
 786         leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 787         leaf.base.obj_type = 0;                 /* unused */
 788         leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;     /* unused */
 789         leaf.base.localization = HAMMER_LOCALIZE_MISC;
 790         leaf.data_len = bytes;
 791
 792         record = hammer_rec_rb_tree_RB_LOOKUP_INFO(&ip->rec_tree, &leaf);
 793         if (record)
 794                 hammer_ref(&record->lock);
 795         return(record);
 796 }
 797
 798 /*
 799  * Reserve blockmap space placemarked with an in-memory record.
 800  *
 801  * This routine is called by the front-end in order to be able to directly
 802  * flush a buffer cache buffer.
 803  */
 804 hammer_record_t
 805 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 806                    int *errorp)
 807 {
 808         hammer_record_t record;
 809         hammer_record_t conflict;
 810         int zone;
 811
 812         /*
 813          * Deal with conflicting in-memory records.  We cannot have multiple
 814          * in-memory records for the same offset without seriously confusing
 815          * the backend, including but not limited to the backend issuing
 816          * delete-create-delete sequences and asserting on the delete_tid
 817          * being the same as the create_tid.
 818          *
 819          * If we encounter a record with the backend interlock set we cannot
 820          * immediately delete it without confusing the backend.
 821          */
 822         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 823                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 824                         conflict->flags |= HAMMER_RECF_WANTED;
 825                         tsleep(conflict, 0, "hmrrc3", 0);
 826                         hammer_rel_mem_record(conflict);
 827                         continue;
 828                 }
 829                 conflict->flags |= HAMMER_RECF_DELETED_FE;
 830                 hammer_rel_mem_record(conflict);
 831         }
 832
 833         /*
 834          * Create a record to cover the direct write.  This is called with
 835          * the related BIO locked so there should be no possible conflict.
 836          *
 837          * The backend is responsible for finalizing the space reserved in
 838          * this record.
 839          *
 840          * XXX bytes not aligned, depend on the reservation code to
 841          * align the reservation.
 842          */
 843         record = hammer_alloc_mem_record(ip, 0);
 844         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 845                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 846         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 847                                                &record->leaf.data_offset,
 848                                                errorp);
 849         if (record->resv == NULL) {
 850                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 851                 hammer_rel_mem_record(record);
 852                 return(NULL);
 853         }
 854         record->type = HAMMER_MEM_RECORD_DATA;
 855         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 856         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 857         record->leaf.base.obj_id = ip->obj_id;
 858         record->leaf.base.key = file_offset + bytes;
 859         record->leaf.base.localization = HAMMER_LOCALIZE_MISC;
 860         record->leaf.data_len = bytes;
 861         record->leaf.data_crc = crc32(data, bytes);
 862
 863         hammer_ref(&record->lock);      /* mem_add eats a reference */
 864         *errorp = hammer_mem_add(record);
 865         KKASSERT(*errorp == 0);
 866
 867         return (record);
 868 }
 869
 870 /*
 871  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 872  * and records in a flushing state are handled by the backend.  The vnops
 873  * setattr code will handle the block containing the truncation point.
 874  *
 875  * Partial blocks are not deleted.
 876  */
 877 int
 878 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
 879 {
 880         struct rec_trunc_info info;
 881
 882         switch(ip->ino_data.obj_type) {
 883         case HAMMER_OBJTYPE_REGFILE:
 884                 info.rec_type = HAMMER_RECTYPE_DATA;
 885                 break;
 886         case HAMMER_OBJTYPE_DBFILE:
 887                 info.rec_type = HAMMER_RECTYPE_DB;
 888                 break;
 889         default:
 890                 return(EINVAL);
 891         }
 892         info.trunc_off = file_size;
 893         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
 894                                    hammer_rec_trunc_callback, &info);
 895         return(0);
 896 }
 897
 898 static int
 899 hammer_rec_trunc_callback(hammer_record_t record, void *data __unused)
 900 {
 901         if (record->flags & HAMMER_RECF_DELETED_FE)
 902                 return(0);
 903         if (record->flush_state == HAMMER_FST_FLUSH)
 904                 return(0);
 905         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
 906         hammer_ref(&record->lock);
 907         record->flags |= HAMMER_RECF_DELETED_FE;
 908         hammer_rel_mem_record(record);
 909         return(0);
 910 }
 911
 912
 913 /*
 914  * Backend code
 915  *
 916  * Sync data from a buffer cache buffer (typically) to the filesystem.  This
 917  * is called via the strategy called from a cached data source.  This code
 918  * is responsible for actually writing a data record out to the disk.
 919  *
 920  * This can only occur non-historically (i.e. 'current' data only).
 921  *
 922  * The file offset must be HAMMER_BUFSIZE aligned but the data length
 923  * can be truncated.  The record (currently) always represents a BUFSIZE
 924  * swath of space whether the data is truncated or not.
 925  */
 926 int
 927 hammer_ip_sync_data(hammer_cursor_t cursor, hammer_inode_t ip,
 928                        int64_t offset, void *data, int bytes)
 929 {
 930         hammer_transaction_t trans = cursor->trans;
 931         struct hammer_btree_leaf_elm elm;
 932         hammer_off_t data_offset;
 933         void *bdata;
 934         int error;
 935         int aligned_bytes;
 936
 937         KKASSERT((offset & HAMMER_BUFMASK) == 0);
 938         KKASSERT(trans->type == HAMMER_TRANS_FLS);
 939         KKASSERT(bytes != 0);
 940
 941         /*
 942          * We don't have to do this but it's probably a good idea to
 943          * align data allocations to 64-byte boundaries for future
 944          * expansion.
 945          */
 946         aligned_bytes = (bytes + 15) & ~15;
 947 retry:
 948         hammer_normalize_cursor(cursor);
 949         cursor->key_beg.localization = HAMMER_LOCALIZE_MISC;
 950         cursor->key_beg.obj_id = ip->obj_id;
 951         cursor->key_beg.key = offset + aligned_bytes;
 952         cursor->key_beg.create_tid = trans->tid;
 953         cursor->key_beg.delete_tid = 0;
 954         cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
 955         cursor->asof = trans->tid;
 956         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
 957         cursor->flags |= HAMMER_CURSOR_INSERT;
 958         cursor->flags |= HAMMER_CURSOR_BACKEND;
 959
 960         /*
 961          * Issue a lookup to position the cursor.
 962          */
 963         error = hammer_btree_lookup(cursor);
 964         if (error == 0) {
 965                 kprintf("hammer_ip_sync_data: duplicate data at "
 966                         "(%lld,%d) tid %016llx\n",
 967                         offset, aligned_bytes, trans->tid);
 968                 hammer_print_btree_elm(&cursor->node->ondisk->
 969                                         elms[cursor->index],
 970                                        HAMMER_BTREE_TYPE_LEAF, cursor->index);
 971                 panic("Duplicate data");
 972                 error = EIO;
 973         }
 974         if (error != ENOENT)
 975                 goto done;
 976
 977         /*
 978          * Allocate our data.  The data buffer is not marked modified (yet)
 979          */
 980         bdata = hammer_alloc_data(trans, aligned_bytes, &data_offset,
 981                                   &cursor->data_buffer, &error);
 982
 983         if (bdata == NULL)
 984                 goto done;
 985
 986         /*
 987          * Fill everything in and insert our B-Tree node.
 988          *
 989          * NOTE: hammer_alloc_data() has already marked the data buffer
 990          * as modified.  If we do it again we will generate unnecessary
 991          * undo elements.
 992          */
 993         elm.base.btype = HAMMER_BTREE_TYPE_RECORD;
 994         elm.base.localization = HAMMER_LOCALIZE_MISC;
 995         elm.base.obj_id = ip->obj_id;
 996         elm.base.key = offset + aligned_bytes;
 997         elm.base.create_tid = trans->tid;
 998         elm.base.delete_tid = 0;
 999         elm.base.rec_type = HAMMER_RECTYPE_DATA;
1000         elm.atime = 0;
1001         elm.data_offset = data_offset;
1002         elm.data_len = aligned_bytes;
1003
1004         /*
1005          * Copy the data to the allocated buffer.  Since we are aligning
1006          * the record size as specified in elm.data_len, make sure to zero
1007          * out any extranious bytes.
1008          */
1009         hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1010         bcopy(data, bdata, bytes);
1011         if (aligned_bytes > bytes)
1012                 bzero((char *)bdata + bytes, aligned_bytes - bytes);
1013         hammer_modify_buffer_done(cursor->data_buffer);
1014         elm.data_crc = crc32(bdata, aligned_bytes);
1015
1016         /*
1017          * Data records can wind up on-disk before the inode itself is
1018          * on-disk.  One must assume data records may be on-disk if either
1019          * HAMMER_INODE_DONDISK or HAMMER_INODE_ONDISK is set
1020          */
1021         ip->flags |= HAMMER_INODE_DONDISK;
1022
1023         error = hammer_btree_insert(cursor, &elm);
1024         if (error == 0)
1025                 goto done;
1026
1027         hammer_blockmap_free(trans, data_offset, aligned_bytes);
1028 done:
1029         if (error == EDEADLK) {
1030                 hammer_done_cursor(cursor);
1031                 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1032                 if (error == 0)
1033                         goto retry;
1034         }
1035         return(error);
1036 }
1037
1038 #if 0
1039
1040 /*
1041  * Backend code which actually performs the write to the media.  This
1042  * routine is typically called from the flusher.  The bio will be disposed
1043  * of (biodone'd) by this routine.
1044  *
1045  * Iterate the related records and mark for deletion.  If existing edge
1046  * records (left and right side) overlap our write they have to be marked
1047  * deleted and new records created, usually referencing a portion of the
1048  * original data.  Then add a record to represent the buffer.
1049  */
1050 int
1051 hammer_dowrite(hammer_cursor_t cursor, hammer_inode_t ip,
1052                off_t file_offset, void *data, int bytes)
1053 {
1054         int error;
1055
1056         KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);
1057
1058         /*
1059          * If the inode is going or gone, just throw away any frontend
1060          * buffers.
1061          */
1062         if (ip->flags & HAMMER_INODE_DELETED)
1063                 return(0);
1064
1065         /*
1066          * Delete any records overlapping our range.  This function will
1067          * (eventually) properly truncate partial overlaps.
1068          */
1069         if (ip->sync_ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1070                 error = hammer_ip_delete_range(cursor, ip, file_offset,
1071                                                file_offset, 0);
1072         } else {
1073                 error = hammer_ip_delete_range(cursor, ip, file_offset,
1074                                                file_offset + bytes - 1, 0);
1075         }
1076
1077         /*
1078          * Add a single record to cover the write.  We can write a record
1079          * with only the actual file data - for example, a small 200 byte
1080          * file does not have to write out a 16K record.
1081          *
1082          * While the data size does not have to be aligned, we still do it
1083          * to reduce fragmentation in a future allocation model.
1084          */
1085         if (error == 0) {
1086                 int limit_size;
1087
1088                 if (ip->sync_ino_data.size - file_offset > bytes) {
1089                             limit_size = bytes;
1090                 } else {
1091                         limit_size = (int)(ip->sync_ino_data.size -
1092                                            file_offset);
1093                         KKASSERT(limit_size >= 0);
1094                 }
1095                 if (limit_size) {
1096                         error = hammer_ip_sync_data(cursor, ip, file_offset,
1097                                                     data, limit_size);
1098                 }
1099         }
1100         if (error)
1101                 Debugger("hammer_dowrite: error");
1102         return(error);
1103 }
1104
1105 #endif
1106
1107 /*
1108  * Backend code.  Sync a record to the media.
1109  */
1110 int
1111 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1112 {
1113         hammer_transaction_t trans = cursor->trans;
1114         int64_t file_offset;
1115         void *bdata;
1116         int error;
1117
1118         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1119         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1120         KKASSERT(record->leaf.base.localization != 0);
1121
1122         /*
1123          * If this is a bulk-data record placemarker there may be an existing
1124          * record on-disk, indicating a data overwrite.  If there is the
1125          * on-disk record must be deleted before we can insert our new record.
1126          *
1127          * We've synthesized this record and do not know what the create_tid
1128          * on-disk is, nor how much data it represents.
1129          *
1130          * Keep in mind that (key) for data records is (base_offset + len),
1131          * not (base_offset).  Also, we only want to get rid of on-disk
1132          * records since we are trying to sync our in-memory record, call
1133          * hammer_ip_delete_range() with truncating set to 1 to make sure
1134          * it skips in-memory records.
1135          *
1136          * It is ok for the lookup to return ENOENT.
1137          */
1138         if (record->type == HAMMER_MEM_RECORD_DATA) {
1139                 file_offset = record->leaf.base.key - record->leaf.data_len;
1140                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1141                 error = hammer_ip_delete_range(
1142                                 cursor, record->ip,
1143                                 file_offset, file_offset + HAMMER_BUFSIZE - 1,
1144                                 1);
1145                 if (error && error != ENOENT)
1146                         goto done;
1147         }
1148
1149         /*
1150          * Setup the cursor.
1151          */
1152         hammer_normalize_cursor(cursor);
1153         cursor->key_beg = record->leaf.base;
1154         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1155         cursor->flags |= HAMMER_CURSOR_BACKEND;
1156         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1157
1158         /*
1159          * Records can wind up on-media before the inode itself is on-media.
1160          * Flag the case.
1161          */
1162         record->ip->flags |= HAMMER_INODE_DONDISK;
1163
1164         /*
1165          * If we are deleting a directory entry an exact match must be
1166          * found on-disk.
1167          */
1168         if (record->type == HAMMER_MEM_RECORD_DEL) {
1169                 error = hammer_btree_lookup(cursor);
1170                 if (error == 0) {
1171                         error = hammer_ip_delete_record(cursor, record->ip,
1172                                                         trans->tid);
1173                         if (error == 0) {
1174                                 record->flags |= HAMMER_RECF_DELETED_FE;
1175                                 record->flags |= HAMMER_RECF_DELETED_BE;
1176                         }
1177                 }
1178                 goto done;
1179         }
1180
1181         /*
1182          * We are inserting.
1183          *
1184          * Issue a lookup to position the cursor and locate the cluster.  The
1185          * target key should not exist.  If we are creating a directory entry
1186          * we may have to iterate the low 32 bits of the key to find an unused
1187          * key.
1188          */
1189         cursor->flags |= HAMMER_CURSOR_INSERT;
1190
1191         for (;;) {
1192                 error = hammer_btree_lookup(cursor);
1193                 if (hammer_debug_inode)
1194                         kprintf("DOINSERT LOOKUP %d\n", error);
1195                 if (error)
1196                         break;
1197                 if (record->leaf.base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
1198                         kprintf("hammer_ip_sync_record: duplicate rec "
1199                                 "at (%016llx)\n", record->leaf.base.key);
1200                         Debugger("duplicate record1");
1201                         error = EIO;
1202                         break;
1203                 }
1204                 if (++trans->hmp->namekey_iterator == 0)
1205                         ++trans->hmp->namekey_iterator;
1206                 record->leaf.base.key &= ~(0xFFFFFFFFLL);
1207                 record->leaf.base.key |= trans->hmp->namekey_iterator;
1208                 cursor->key_beg.key = record->leaf.base.key;
1209         }
1210 #if 0
1211         if (record->type == HAMMER_MEM_RECORD_DATA)
1212                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1213                         record->leaf.base.key - record->leaf.data_len,
1214                         record->leaf.data_offset, error);
1215 #endif
1216
1217
1218         if (error != ENOENT)
1219                 goto done;
1220
1221         /*
1222          * Allocate the record and data.  The result buffers will be
1223          * marked as being modified and further calls to
1224          * hammer_modify_buffer() will result in unneeded UNDO records.
1225          *
1226          * Support zero-fill records (data == NULL and data_len != 0)
1227          */
1228         if (record->type == HAMMER_MEM_RECORD_DATA) {
1229                 /*
1230                  * The data portion of a bulk-data record has already been
1231                  * committed to disk, we need only adjust the layer2
1232                  * statistics in the same transaction as our B-Tree insert.
1233                  */
1234                 KKASSERT(record->leaf.data_offset != 0);
1235                 hammer_blockmap_free(trans, record->leaf.data_offset,
1236                                      -record->leaf.data_len);
1237                 error = 0;
1238         } else if (record->data && record->leaf.data_len) {
1239                 /*
1240                  * Wholely cached record, with data.  Allocate the data.
1241                  */
1242                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1243                                           &record->leaf.data_offset,
1244                                           &cursor->data_buffer, &error);
1245                 if (bdata == NULL)
1246                         goto done;
1247                 record->leaf.data_crc = crc32(record->data,
1248                                               record->leaf.data_len);
1249                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1250                 bcopy(record->data, bdata, record->leaf.data_len);
1251                 hammer_modify_buffer_done(cursor->data_buffer);
1252         } else {
1253                 /*
1254                  * Wholely cached record, without data.
1255                  */
1256                 record->leaf.data_offset = 0;
1257                 record->leaf.data_crc = 0;
1258         }
1259
1260         error = hammer_btree_insert(cursor, &record->leaf);
1261         if (hammer_debug_inode && error)
1262                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1263
1264         /*
1265          * Our record is on-disk, normally mark the in-memory version as
1266          * deleted.  If the record represented a directory deletion but
1267          * we had to sync a valid directory entry to disk we must convert
1268          * the record to a covering delete so the frontend does not have
1269          * visibility on the synced entry.
1270          */
1271         if (error == 0) {
1272                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1273                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1274                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1275                         record->type = HAMMER_MEM_RECORD_DEL;
1276                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1277                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1278                         /* hammer_flush_record_done takes care of the rest */
1279                 } else {
1280                         record->flags |= HAMMER_RECF_DELETED_FE;
1281                         record->flags |= HAMMER_RECF_DELETED_BE;
1282                 }
1283         } else {
1284                 if (record->leaf.data_offset) {
1285                         hammer_blockmap_free(trans, record->leaf.data_offset,
1286                                              record->leaf.data_len);
1287                 }
1288         }
1289
1290 done:
1291         return(error);
1292 }
1293
1294 /*
1295  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1296  * entry's key is used to deal with hash collisions in the upper 32 bits.
1297  * A unique 64 bit key is generated in-memory and may be regenerated a
1298  * second time when the directory record is flushed to the on-disk B-Tree.
1299  *
1300  * A referenced record is passed to this function.  This function
1301  * eats the reference.  If an error occurs the record will be deleted.
1302  *
1303  * A copy of the temporary record->data pointer provided by the caller
1304  * will be made.
1305  */
1306 static
1307 int
1308 hammer_mem_add(hammer_record_t record)
1309 {
1310         hammer_mount_t hmp = record->ip->hmp;
1311
1312         /*
1313          * Make a private copy of record->data
1314          */
1315         if (record->data)
1316                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1317
1318         /*
1319          * Insert into the RB tree, find an unused iterator if this is
1320          * a directory entry.
1321          */
1322         while (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1323                 if (record->leaf.base.rec_type != HAMMER_RECTYPE_DIRENTRY){
1324                         record->flags |= HAMMER_RECF_DELETED_FE;
1325                         hammer_rel_mem_record(record);
1326                         return (EEXIST);
1327                 }
1328                 if (++hmp->namekey_iterator == 0)
1329                         ++hmp->namekey_iterator;
1330                 record->leaf.base.key &= ~(0xFFFFFFFFLL);
1331                 record->leaf.base.key |= hmp->namekey_iterator;
1332         }
1333         ++hmp->rsv_recs;
1334         ++record->ip->rsv_recs;
1335         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1336         record->flags |= HAMMER_RECF_ONRBTREE;
1337         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1338         hammer_rel_mem_record(record);
1339         return(0);
1340 }
1341
1342 /************************************************************************
1343  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1344  ************************************************************************
1345  *
1346  * These functions augment the B-Tree scanning functions in hammer_btree.c
1347  * by merging in-memory records with on-disk records.
1348  */
1349
1350 /*
1351  * Locate a particular record either in-memory or on-disk.
1352  *
1353  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1354  * NOT be called to iterate results.
1355  */
1356 int
1357 hammer_ip_lookup(hammer_cursor_t cursor)
1358 {
1359         int error;
1360
1361         /*
1362          * If the element is in-memory return it without searching the
1363          * on-disk B-Tree
1364          */
1365         KKASSERT(cursor->ip);
1366         error = hammer_mem_lookup(cursor);
1367         if (error == 0) {
1368                 cursor->leaf = &cursor->iprec->leaf;
1369                 return(error);
1370         }
1371         if (error != ENOENT)
1372                 return(error);
1373
1374         /*
1375          * If the inode has on-disk components search the on-disk B-Tree.
1376          */
1377         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1378                 return(error);
1379         error = hammer_btree_lookup(cursor);
1380         if (error == 0)
1381                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1382         return(error);
1383 }
1384
1385 /*
1386  * Locate the first record within the cursor's key_beg/key_end range,
1387  * restricted to a particular inode.  0 is returned on success, ENOENT
1388  * if no records matched the requested range, or some other error.
1389  *
1390  * When 0 is returned hammer_ip_next() may be used to iterate additional
1391  * records within the requested range.
1392  *
1393  * This function can return EDEADLK, requiring the caller to terminate
1394  * the cursor and try again.
1395  */
1396 int
1397 hammer_ip_first(hammer_cursor_t cursor)
1398 {
1399         hammer_inode_t ip = cursor->ip;
1400         int error;
1401
1402         KKASSERT(ip != NULL);
1403
1404         /*
1405          * Clean up fields and setup for merged scan
1406          */
1407         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1408         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1409         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1410         if (cursor->iprec) {
1411                 hammer_rel_mem_record(cursor->iprec);
1412                 cursor->iprec = NULL;
1413         }
1414
1415         /*
1416          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1417          * exact lookup so if we get ENOENT we have to call the iterate
1418          * function to validate the first record after the begin key.
1419          *
1420          * The ATEDISK flag is used by hammer_btree_iterate to determine
1421          * whether it must index forwards or not.  It is also used here
1422          * to select the next record from in-memory or on-disk.
1423          *
1424          * EDEADLK can only occur if the lookup hit an empty internal
1425          * element and couldn't delete it.  Since this could only occur
1426          * in-range, we can just iterate from the failure point.
1427          */
1428         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1429                 error = hammer_btree_lookup(cursor);
1430                 if (error == ENOENT || error == EDEADLK) {
1431                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1432                         if (hammer_debug_general & 0x2000)
1433                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1434                         error = hammer_btree_iterate(cursor);
1435                 }
1436                 if (error && error != ENOENT)
1437                         return(error);
1438                 if (error == 0) {
1439                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1440                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1441                 } else {
1442                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1443                 }
1444         }
1445
1446         /*
1447          * Search the in-memory record list (Red-Black tree).  Unlike the
1448          * B-Tree search, mem_first checks for records in the range.
1449          */
1450         error = hammer_mem_first(cursor);
1451         if (error && error != ENOENT)
1452                 return(error);
1453         if (error == 0) {
1454                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1455                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1456                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1457                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1458         }
1459
1460         /*
1461          * This will return the first matching record.
1462          */
1463         return(hammer_ip_next(cursor));
1464 }
1465
1466 /*
1467  * Retrieve the next record in a merged iteration within the bounds of the
1468  * cursor.  This call may be made multiple times after the cursor has been
1469  * initially searched with hammer_ip_first().
1470  *
1471  * 0 is returned on success, ENOENT if no further records match the
1472  * requested range, or some other error code is returned.
1473  */
1474 int
1475 hammer_ip_next(hammer_cursor_t cursor)
1476 {
1477         hammer_btree_elm_t elm;
1478         hammer_record_t rec, save;
1479         int error;
1480         int r;
1481
1482 next_btree:
1483         /*
1484          * Load the current on-disk and in-memory record.  If we ate any
1485          * records we have to get the next one.
1486          *
1487          * If we deleted the last on-disk record we had scanned ATEDISK will
1488          * be clear and DELBTREE will be set, forcing a call to iterate. The
1489          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1490          * element.  If ATEDISK is set, iterate will skip the 'current'
1491          * element.
1492          *
1493          * Get the next on-disk record
1494          */
1495         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1496                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1497                         error = hammer_btree_iterate(cursor);
1498                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1499                         if (error == 0)
1500                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1501                         else
1502                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1503                                                  HAMMER_CURSOR_ATEDISK;
1504                 }
1505         }
1506
1507 next_memory:
1508         /*
1509          * Get the next in-memory record.  The record can be ripped out
1510          * of the RB tree so we maintain a scan_info structure to track
1511          * the next node.
1512          *
1513          * hammer_rec_scan_cmp:  Is the record still in our general range,
1514          *                       (non-inclusive of snapshot exclusions)?
1515          * hammer_rec_scan_callback: Is the record in our snapshot?
1516          */
1517         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1518                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1519                         save = cursor->iprec;
1520                         cursor->iprec = NULL;
1521                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1522                         while (rec) {
1523                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1524                                         break;
1525                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1526                                         break;
1527                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1528                         }
1529                         if (save)
1530                                 hammer_rel_mem_record(save);
1531                         if (cursor->iprec) {
1532                                 KKASSERT(cursor->iprec == rec);
1533                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1534                         } else {
1535                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1536                         }
1537                 }
1538         }
1539
1540         /*
1541          * The memory record may have become stale while being held in
1542          * cursor->iprec.  We are interlocked against the backend on
1543          * with regards to B-Tree entries.
1544          */
1545         if ((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0) {
1546                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0) {
1547                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1548                         goto next_memory;
1549                 }
1550         }
1551
1552         /*
1553          * Extract either the disk or memory record depending on their
1554          * relative position.
1555          */
1556         error = 0;
1557         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1558         case 0:
1559                 /*
1560                  * Both entries valid.   Compare the entries and nominally
1561                  * return the first one in the sort order.  Numerous cases
1562                  * require special attention, however.
1563                  */
1564                 elm = &cursor->node->ondisk->elms[cursor->index];
1565                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1566
1567                 /*
1568                  * If the two entries differ only by their key (-2/2) or
1569                  * create_tid (-1/1), and are DATA records, we may have a
1570                  * nominal match.  We have to calculate the base file
1571                  * offset of the data.
1572                  */
1573                 if (r <= 2 && r >= -2 && r != 0 &&
1574                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1575                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1576                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1577                         int64_t base2 = cursor->iprec->leaf.base.key -
1578                                         cursor->iprec->leaf.data_len;
1579                         if (base1 == base2) {
1580                                 kprintf("G");
1581                                 r = 0;
1582                         }
1583                 }
1584
1585                 if (r < 0) {
1586                         error = hammer_btree_extract(cursor,
1587                                                      HAMMER_CURSOR_GET_LEAF);
1588                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1589                         break;
1590                 }
1591
1592                 /*
1593                  * If the entries match exactly the memory entry is either
1594                  * an on-disk directory entry deletion or a bulk data
1595                  * overwrite.  If it is a directory entry deletion we eat
1596                  * both entries.
1597                  *
1598                  * For the bulk-data overwrite case it is possible to have
1599                  * visibility into both, which simply means the syncer
1600                  * hasn't gotten around to doing the delete+insert sequence
1601                  * on the B-Tree.  Use the memory entry and throw away the
1602                  * on-disk entry.
1603                  *
1604                  * If the in-memory record is not either of these we
1605                  * probably caught the syncer while it was syncing it to
1606                  * the media.  Since we hold a shared lock on the cursor,
1607                  * the in-memory record had better be marked deleted at
1608                  * this point.
1609                  */
1610                 if (r == 0) {
1611                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1612                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1613                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1614                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1615                                         goto next_btree;
1616                                 }
1617                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1618                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1619                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1620                                 }
1621                                 /* fall through to memory entry */
1622                         } else {
1623                                 panic("hammer_ip_next: duplicate mem/b-tree entry");
1624                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1625                                 goto next_memory;
1626                         }
1627                 }
1628                 /* fall through to the memory entry */
1629         case HAMMER_CURSOR_ATEDISK:
1630                 /*
1631                  * Only the memory entry is valid.
1632                  */
1633                 cursor->leaf = &cursor->iprec->leaf;
1634                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1635
1636                 /*
1637                  * If the memory entry is an on-disk deletion we should have
1638                  * also had found a B-Tree record.  If the backend beat us
1639                  * to it it would have interlocked the cursor and we should
1640                  * have seen the in-memory record marked DELETED_FE.
1641                  */
1642                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1643                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1644                         panic("hammer_ip_next: del-on-disk with no b-tree entry");
1645                 }
1646                 break;
1647         case HAMMER_CURSOR_ATEMEM:
1648                 /*
1649                  * Only the disk entry is valid
1650                  */
1651                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1652                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1653                 break;
1654         default:
1655                 /*
1656                  * Neither entry is valid
1657                  *
1658                  * XXX error not set properly
1659                  */
1660                 cursor->leaf = NULL;
1661                 error = ENOENT;
1662                 break;
1663         }
1664         return(error);
1665 }
1666
1667 /*
1668  * Resolve the cursor->data pointer for the current cursor position in
1669  * a merged iteration.
1670  */
1671 int
1672 hammer_ip_resolve_data(hammer_cursor_t cursor)
1673 {
1674         hammer_record_t record;
1675         int error;
1676
1677         if (hammer_cursor_inmem(cursor)) {
1678                 /*
1679                  * The data associated with an in-memory record is usually
1680                  * kmalloced, but reserve-ahead data records will have an
1681                  * on-disk reference.
1682                  *
1683                  * NOTE: Reserve-ahead data records must be handled in the
1684                  * context of the related high level buffer cache buffer
1685                  * to interlock against async writes.
1686                  */
1687                 record = cursor->iprec;
1688                 cursor->data = record->data;
1689                 error = 0;
1690                 if (cursor->data == NULL) {
1691                         KKASSERT(record->leaf.base.rec_type ==
1692                                  HAMMER_RECTYPE_DATA);
1693                         cursor->data = hammer_bread(cursor->trans->hmp,
1694                                                     record->leaf.data_offset,
1695                                                     &error,
1696                                                     &cursor->data_buffer);
1697                 }
1698         } else {
1699                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1700                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1701         }
1702         return(error);
1703 }
1704
1705 /*
1706  * Backend truncation / record replacement - delete records in range.
1707  *
1708  * Delete all records within the specified range for inode ip.  In-memory
1709  * records still associated with the frontend are ignored.
1710  *
1711  * NOTE: An unaligned range will cause new records to be added to cover
1712  * the edge cases. (XXX not implemented yet).
1713  *
1714  * NOTE: Replacement via reservations (see hammer_ip_sync_record_cursor())
1715  * also do not deal with unaligned ranges.
1716  *
1717  * NOTE: ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1718  *
1719  * NOTE: Record keys for regular file data have to be special-cased since
1720  * they indicate the end of the range (key = base + bytes).
1721  */
1722 int
1723 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1724                        int64_t ran_beg, int64_t ran_end, int truncating)
1725 {
1726         hammer_transaction_t trans = cursor->trans;
1727         hammer_btree_leaf_elm_t leaf;
1728         int error;
1729         int64_t off;
1730
1731 #if 0
1732         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1733 #endif
1734
1735         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1736 retry:
1737         hammer_normalize_cursor(cursor);
1738         cursor->key_beg.localization = HAMMER_LOCALIZE_MISC;
1739         cursor->key_beg.obj_id = ip->obj_id;
1740         cursor->key_beg.create_tid = 0;
1741         cursor->key_beg.delete_tid = 0;
1742         cursor->key_beg.obj_type = 0;
1743         cursor->asof = ip->obj_asof;
1744         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1745         cursor->flags |= HAMMER_CURSOR_ASOF;
1746         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1747         cursor->flags |= HAMMER_CURSOR_BACKEND;
1748
1749         cursor->key_end = cursor->key_beg;
1750         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1751                 cursor->key_beg.key = ran_beg;
1752                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1753                 cursor->key_end.rec_type = HAMMER_RECTYPE_DB;
1754                 cursor->key_end.key = ran_end;
1755         } else {
1756                 /*
1757                  * The key in the B-Tree is (base+bytes), so the first possible
1758                  * matching key is ran_beg + 1.
1759                  */
1760                 int64_t tmp64;
1761
1762                 cursor->key_beg.key = ran_beg + 1;
1763                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1764                 cursor->key_end.rec_type = HAMMER_RECTYPE_DATA;
1765
1766                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1767                 if (tmp64 < ran_end)
1768                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1769                 else
1770                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1771         }
1772         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1773
1774         error = hammer_ip_first(cursor);
1775
1776         /*
1777          * Iterate through matching records and mark them as deleted.
1778          */
1779         while (error == 0) {
1780                 leaf = cursor->leaf;
1781
1782                 KKASSERT(leaf->base.delete_tid == 0);
1783
1784                 /*
1785                  * There may be overlap cases for regular file data.  Also
1786                  * remember the key for a regular file record is (base + len),
1787                  * NOT (base).
1788                  */
1789                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1790                         off = leaf->base.key - leaf->data_len;
1791                         /*
1792                          * Check the left edge case.  We currently do not
1793                          * split existing records.
1794                          */
1795                         if (off < ran_beg) {
1796                                 panic("hammer left edge case %016llx %d\n",
1797                                         leaf->base.key, leaf->data_len);
1798                         }
1799
1800                         /*
1801                          * Check the right edge case.  Note that the
1802                          * record can be completely out of bounds, which
1803                          * terminates the search.
1804                          *
1805                          * base->key is exclusive of the right edge while
1806                          * ran_end is inclusive of the right edge.  The
1807                          * (key - data_len) left boundary is inclusive.
1808                          *
1809                          * XXX theory-check this test at some point, are
1810                          * we missing a + 1 somewhere?  Note that ran_end
1811                          * could overflow.
1812                          */
1813                         if (leaf->base.key - 1 > ran_end) {
1814                                 if (leaf->base.key - leaf->data_len > ran_end)
1815                                         break;
1816                                 panic("hammer right edge case\n");
1817                         }
1818                 }
1819
1820                 /*
1821                  * Delete the record.  When truncating we do not delete
1822                  * in-memory (data) records because they represent data
1823                  * written after the truncation.
1824                  *
1825                  * This will also physically destroy the B-Tree entry and
1826                  * data if the retention policy dictates.  The function
1827                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1828                  * uses to perform a fixup.
1829                  */
1830                 if (truncating == 0 || hammer_cursor_ondisk(cursor))
1831                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1832                 if (error)
1833                         break;
1834                 error = hammer_ip_next(cursor);
1835         }
1836         if (error == EDEADLK) {
1837                 hammer_done_cursor(cursor);
1838                 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1839                 if (error == 0)
1840                         goto retry;
1841         }
1842         if (error == ENOENT)
1843                 error = 0;
1844         return(error);
1845 }
1846
1847 /*
1848  * Backend truncation - delete all records.
1849  *
1850  * Delete all user records associated with an inode except the inode record
1851  * itself.  Directory entries are not deleted (they must be properly disposed
1852  * of or nlinks would get upset).
1853  */
1854 int
1855 hammer_ip_delete_range_all(hammer_cursor_t cursor, hammer_inode_t ip,
1856                            int *countp)
1857 {
1858         hammer_transaction_t trans = cursor->trans;
1859         hammer_btree_leaf_elm_t leaf;
1860         int error;
1861
1862         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1863 retry:
1864         hammer_normalize_cursor(cursor);
1865         cursor->key_beg.localization = HAMMER_LOCALIZE_MISC;
1866         cursor->key_beg.obj_id = ip->obj_id;
1867         cursor->key_beg.create_tid = 0;
1868         cursor->key_beg.delete_tid = 0;
1869         cursor->key_beg.obj_type = 0;
1870         cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1871         cursor->key_beg.key = HAMMER_MIN_KEY;
1872
1873         cursor->key_end = cursor->key_beg;
1874         cursor->key_end.rec_type = 0xFFFF;
1875         cursor->key_end.key = HAMMER_MAX_KEY;
1876
1877         cursor->asof = ip->obj_asof;
1878         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1879         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1880         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1881         cursor->flags |= HAMMER_CURSOR_BACKEND;
1882
1883         error = hammer_ip_first(cursor);
1884
1885         /*
1886          * Iterate through matching records and mark them as deleted.
1887          */
1888         while (error == 0) {
1889                 leaf = cursor->leaf;
1890
1891                 KKASSERT(leaf->base.delete_tid == 0);
1892
1893                 /*
1894                  * Mark the record and B-Tree entry as deleted.  This will
1895                  * also physically delete the B-Tree entry, record, and
1896                  * data if the retention policy dictates.  The function
1897                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1898                  * uses to perform a fixup.
1899                  *
1900                  * Directory entries (and delete-on-disk directory entries)
1901                  * must be synced and cannot be deleted.
1902                  */
1903                 if (leaf->base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
1904                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1905                         ++*countp;
1906                 }
1907                 if (error)
1908                         break;
1909                 error = hammer_ip_next(cursor);
1910         }
1911         if (error == EDEADLK) {
1912                 hammer_done_cursor(cursor);
1913                 error = hammer_init_cursor(trans, cursor, &ip->cache[0], ip);
1914                 if (error == 0)
1915                         goto retry;
1916         }
1917         if (error == ENOENT)
1918                 error = 0;
1919         return(error);
1920 }
1921
1922 /*
1923  * Delete the record at the current cursor.  On success the cursor will
1924  * be positioned appropriately for an iteration but may no longer be at
1925  * a leaf node.
1926  *
1927  * This routine is only called from the backend.
1928  *
1929  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1930  * cursor and retry.
1931  */
1932 int
1933 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
1934                         hammer_tid_t tid)
1935 {
1936         hammer_btree_elm_t elm;
1937         hammer_mount_t hmp;
1938         int error;
1939         int dodelete;
1940
1941         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
1942         KKASSERT(tid != 0);
1943
1944         /*
1945          * In-memory (unsynchronized) records can simply be freed.  This
1946          * only occurs in range iterations since all other records are
1947          * individually synchronized.  Thus there should be no confusion with
1948          * the interlock.
1949          */
1950         if (hammer_cursor_inmem(cursor)) {
1951                 KKASSERT((cursor->iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
1952                 cursor->iprec->flags |= HAMMER_RECF_DELETED_FE;
1953                 cursor->iprec->flags |= HAMMER_RECF_DELETED_BE;
1954                 return(0);
1955         }
1956
1957         /*
1958          * On-disk records are marked as deleted by updating their delete_tid.
1959          * This does not effect their position in the B-Tree (which is based
1960          * on their create_tid).
1961          */
1962         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1963         elm = NULL;
1964         hmp = cursor->node->hmp;
1965
1966         /*
1967          * If we were mounted with the nohistory option, we physically
1968          * delete the record.
1969          */
1970         dodelete = hammer_nohistory(ip);
1971
1972         if (error == 0) {
1973                 error = hammer_cursor_upgrade(cursor);
1974                 if (error == 0) {
1975                         elm = &cursor->node->ondisk->elms[cursor->index];
1976                         hammer_modify_node(cursor->trans, cursor->node,
1977                                            &elm->leaf.base.delete_tid,
1978                                            sizeof(elm->leaf.base.delete_tid));
1979                         elm->leaf.base.delete_tid = tid;
1980                         hammer_modify_node_done(cursor->node);
1981
1982                         /*
1983                          * An on-disk record cannot have the same delete_tid
1984                          * as its create_tid.  In a chain of record updates
1985                          * this could result in a duplicate record.
1986                          */
1987                         KKASSERT(elm->leaf.base.delete_tid != elm->leaf.base.create_tid);
1988                 }
1989         }
1990
1991         if (error == 0 && dodelete) {
1992                 error = hammer_delete_at_cursor(cursor, NULL);
1993                 if (error) {
1994                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1995                         error = 0;
1996                 }
1997         }
1998         return(error);
1999 }
2000
2001 int
2002 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
2003 {
2004         hammer_btree_elm_t elm;
2005         hammer_off_t data_offset;
2006         int32_t data_len;
2007         u_int16_t rec_type;
2008         int error;
2009
2010         elm = &cursor->node->ondisk->elms[cursor->index];
2011         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2012
2013         data_offset = elm->leaf.data_offset;
2014         data_len = elm->leaf.data_len;
2015         rec_type = elm->leaf.base.rec_type;
2016
2017         error = hammer_btree_delete(cursor);
2018         if (error == 0) {
2019                 /*
2020                  * This forces a fixup for the iteration because
2021                  * the cursor is now either sitting at the 'next'
2022                  * element or sitting at the end of a leaf.
2023                  */
2024                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2025                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
2026                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2027                 }
2028         }
2029         if (error == 0) {
2030                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2031                 case HAMMER_ZONE_LARGE_DATA:
2032                 case HAMMER_ZONE_SMALL_DATA:
2033                         hammer_blockmap_free(cursor->trans,
2034                                              data_offset, data_len);
2035                         break;
2036                 default:
2037                         break;
2038                 }
2039         }
2040         return (error);
2041 }
2042
2043 /*
2044  * Determine whether we can remove a directory.  This routine checks whether
2045  * a directory is empty or not and enforces flush connectivity.
2046  *
2047  * Flush connectivity requires that we block if the target directory is
2048  * currently flushing, otherwise it may not end up in the same flush group.
2049  *
2050  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2051  */
2052 int
2053 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2054 {
2055         struct hammer_cursor cursor;
2056         int error;
2057
2058         /*
2059          * Check directory empty
2060          */
2061         hammer_init_cursor(trans, &cursor, &ip->cache[0], ip);
2062
2063         cursor.key_beg.localization = HAMMER_LOCALIZE_MISC;
2064         cursor.key_beg.obj_id = ip->obj_id;
2065         cursor.key_beg.create_tid = 0;
2066         cursor.key_beg.delete_tid = 0;
2067         cursor.key_beg.obj_type = 0;
2068         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2069         cursor.key_beg.key = HAMMER_MIN_KEY;
2070
2071         cursor.key_end = cursor.key_beg;
2072         cursor.key_end.rec_type = 0xFFFF;
2073         cursor.key_end.key = HAMMER_MAX_KEY;
2074
2075         cursor.asof = ip->obj_asof;
2076         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2077
2078         error = hammer_ip_first(&cursor);
2079         if (error == ENOENT)
2080                 error = 0;
2081         else if (error == 0)
2082                 error = ENOTEMPTY;
2083         hammer_done_cursor(&cursor);
2084         return(error);
2085 }
2086