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 1.44 2008/04/26 08:02:17 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_add(hammer_transaction_t trans, hammer_record_t record);
  40 static int hammer_mem_lookup(hammer_cursor_t cursor, hammer_inode_t ip);
  41 static int hammer_mem_first(hammer_cursor_t cursor, hammer_inode_t ip);
  42
  43 /*
  44  * Red-black tree support.
  45  */
  46 static int
  47 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  48 {
  49         if (rec1->rec.base.base.rec_type < rec2->rec.base.base.rec_type)
  50                 return(-1);
  51         if (rec1->rec.base.base.rec_type > rec2->rec.base.base.rec_type)
  52                 return(1);
  53
  54         if (rec1->rec.base.base.key < rec2->rec.base.base.key)
  55                 return(-1);
  56         if (rec1->rec.base.base.key > rec2->rec.base.base.key)
  57                 return(1);
  58
  59         if (rec1->rec.base.base.create_tid == 0) {
  60                 if (rec2->rec.base.base.create_tid == 0)
  61                         return(0);
  62                 return(1);
  63         }
  64         if (rec2->rec.base.base.create_tid == 0)
  65                 return(-1);
  66
  67         if (rec1->rec.base.base.create_tid < rec2->rec.base.base.create_tid)
  68                 return(-1);
  69         if (rec1->rec.base.base.create_tid > rec2->rec.base.base.create_tid)
  70                 return(1);
  71
  72         /*
  73          * Never match against a deleted item.
  74          */
  75         if (rec1->flags & HAMMER_RECF_DELETED_FE)
  76                 return(1);
  77         if (rec2->flags & HAMMER_RECF_DELETED_FE)
  78                 return(-1);
  79
  80         return(0);
  81 }
  82
  83 static int
  84 hammer_rec_compare(hammer_base_elm_t info, hammer_record_t rec)
  85 {
  86         if (info->rec_type < rec->rec.base.base.rec_type)
  87                 return(-3);
  88         if (info->rec_type > rec->rec.base.base.rec_type)
  89                 return(3);
  90
  91         if (info->key < rec->rec.base.base.key)
  92                 return(-2);
  93         if (info->key > rec->rec.base.base.key)
  94                 return(2);
  95
  96         if (info->create_tid == 0) {
  97                 if (rec->rec.base.base.create_tid == 0)
  98                         return(0);
  99                 return(1);
 100         }
 101         if (rec->rec.base.base.create_tid == 0)
 102                 return(-1);
 103         if (info->create_tid < rec->rec.base.base.create_tid)
 104                 return(-1);
 105         if (info->create_tid > rec->rec.base.base.create_tid)
 106                 return(1);
 107         return(0);
 108 }
 109
 110 /*
 111  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 112  * is reversed so the comparison result has to be negated.  key_beg and
 113  * key_end are both range-inclusive.
 114  *
 115  * The creation timestamp can cause hammer_rec_compare() to return -1 or +1.
 116  * These do not stop the scan.
 117  *
 118  * Localized deletions are not cached in-memory.
 119  */
 120 static
 121 int
 122 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 123 {
 124         hammer_cursor_t cursor = data;
 125         int r;
 126
 127         r = hammer_rec_compare(&cursor->key_beg, rec);
 128         if (r > 1)
 129                 return(-1);
 130         r = hammer_rec_compare(&cursor->key_end, rec);
 131         if (r < -1)
 132                 return(1);
 133         return(0);
 134 }
 135
 136 /*
 137  * This compare function is used when simply looking up key_beg.
 138  */
 139 static
 140 int
 141 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 142 {
 143         hammer_cursor_t cursor = data;
 144         int r;
 145
 146         r = hammer_rec_compare(&cursor->key_beg, rec);
 147         if (r > 1)
 148                 return(-1);
 149         if (r < -1)
 150                 return(1);
 151         return(0);
 152 }
 153
 154 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 155 RB_GENERATE_XLOOKUP(hammer_rec_rb_tree, INFO, hammer_record, rb_node,
 156                     hammer_rec_compare, hammer_base_elm_t);
 157
 158 /*
 159  * Allocate a record for the caller to finish filling in.  The record is
 160  * returned referenced.
 161  */
 162 hammer_record_t
 163 hammer_alloc_mem_record(hammer_inode_t ip)
 164 {
 165         hammer_record_t record;
 166
 167         ++hammer_count_records;
 168         record = kmalloc(sizeof(*record), M_HAMMER, M_WAITOK|M_ZERO);
 169         record->state = HAMMER_FST_IDLE;
 170         record->ip = ip;
 171         record->rec.base.base.btype = HAMMER_BTREE_TYPE_RECORD;
 172         TAILQ_INIT(&record->depend_list);
 173         hammer_ref(&record->lock);
 174         return (record);
 175 }
 176
 177 void
 178 hammer_wait_mem_record(hammer_record_t record)
 179 {
 180         while (record->state == HAMMER_FST_FLUSH) {
 181                 record->flags |= HAMMER_RECF_WANTED;
 182                 tsleep(record, 0, "hmrrc2", 0);
 183         }
 184 }
 185
 186 /*
 187  * Called from the backend, hammer_inode.c, when a record has been
 188  * flushed to disk.
 189  *
 190  * The backend has likely marked this record for deletion as well.
 191  */
 192 void
 193 hammer_flush_record_done(hammer_record_t record)
 194 {
 195         KKASSERT(record->state == HAMMER_FST_FLUSH);
 196         record->state = HAMMER_FST_IDLE;
 197         if (record->flags & HAMMER_RECF_WANTED) {
 198                 record->flags &= ~HAMMER_RECF_WANTED;
 199                 wakeup(record);
 200         }
 201         hammer_rel_mem_record(record);
 202 }
 203
 204 /*
 205  * Destroy a memory record.  The record is marked deleted and will
 206  * be freed when the last reference goes away.   Any dependancies are
 207  * also destroyed.
 208  */
 209 void
 210 hammer_delete_mem_record(hammer_record_t record)
 211 {
 212         hammer_depend_t depend;
 213
 214         record->flags |= HAMMER_RECF_DELETED_FE;
 215         while ((depend = TAILQ_FIRST(&record->depend_list)) != NULL) {
 216                 TAILQ_REMOVE(&record->depend_list, depend, rec_entry);
 217                 TAILQ_REMOVE(&depend->ip->depend_list, depend, ip_entry);
 218                 --depend->ip->depend_count;
 219                 /* NOTE: inode is not flushed */
 220                 hammer_rel_inode(depend->ip, 0);
 221                 hammer_unref(&record->lock);
 222                 KKASSERT(record->lock.refs > 0);
 223                 kfree(depend, M_HAMMER);
 224         }
 225 }
 226
 227 /*
 228  * Release a memory record.  Records marked for deletion are immediately
 229  * removed from the RB-Tree but otherwise left intact until the last ref
 230  * goes away.
 231  */
 232 void
 233 hammer_rel_mem_record(struct hammer_record *record)
 234 {
 235         hammer_unref(&record->lock);
 236
 237         if (record->flags & HAMMER_RECF_DELETED_FE) {
 238                 if (record->lock.refs == 0) {
 239                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 240                                 RB_REMOVE(hammer_rec_rb_tree,
 241                                           &record->ip->rec_tree,
 242                                           record);
 243                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 244                         }
 245                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 246                                 --hammer_count_record_datas;
 247                                 kfree(record->data, M_HAMMER);
 248                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 249                         }
 250                         record->data = NULL;
 251                         --hammer_count_records;
 252                         kfree(record, M_HAMMER);
 253                         return;
 254                 }
 255         }
 256
 257         /*
 258          * If someone wanted the record wake them up.
 259          */
 260         if (record->flags & HAMMER_RECF_WANTED) {
 261                 record->flags &= ~HAMMER_RECF_WANTED;
 262                 wakeup(record);
 263         }
 264 }
 265
 266 /*
 267  * The deletion state of a record will appear different to the backend
 268  * then it does to the frontend.
 269  */
 270 static __inline
 271 int
 272 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t rec)
 273 {
 274         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 275                 if (rec->flags & HAMMER_RECF_DELETED_BE)
 276                         return(0);
 277         } else {
 278                 if (rec->flags & HAMMER_RECF_DELETED_FE)
 279                         return(0);
 280         }
 281         return(1);
 282 }
 283
 284 /*
 285  * This callback is used as part of the RB_SCAN function for in-memory
 286  * records.  We terminate it (return -1) as soon as we get a match.
 287  *
 288  * This routine is used by frontend code.
 289  *
 290  * The primary compare code does not account for ASOF lookups.  This
 291  * code handles that case as well as a few others.
 292  */
 293 static
 294 int
 295 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 296 {
 297         hammer_cursor_t cursor = data;
 298
 299         /*
 300          * We terminate on success, so this should be NULL on entry.
 301          */
 302         KKASSERT(cursor->iprec == NULL);
 303
 304         /*
 305          * Skip if the record was marked deleted.
 306          */
 307         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 308                 return(0);
 309
 310         /*
 311          * Skip if not visible due to our as-of TID
 312          */
 313         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 314                 if (cursor->asof < rec->rec.base.base.create_tid)
 315                         return(0);
 316                 if (rec->rec.base.base.delete_tid &&
 317                     cursor->asof >= rec->rec.base.base.delete_tid) {
 318                         return(0);
 319                 }
 320         }
 321
 322         /*
 323          * If the record is queued to the flusher we have to block until
 324          * it isn't.  Otherwise we may see duplication between our memory
 325          * cache and the media.
 326          */
 327         hammer_ref(&rec->lock);
 328
 329 #warning "This deadlocks"
 330 #if 0
 331         if (rec->state == HAMMER_FST_FLUSH)
 332                 hammer_wait_mem_record(rec);
 333 #endif
 334
 335         /*
 336          * The record may have been deleted while we were blocked.
 337          */
 338         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 339                 hammer_rel_mem_record(rec);
 340                 return(0);
 341         }
 342
 343         /*
 344          * Set the matching record and stop the scan.
 345          */
 346         cursor->iprec = rec;
 347         return(-1);
 348 }
 349
 350
 351 /*
 352  * Lookup an in-memory record given the key specified in the cursor.  Works
 353  * just like hammer_btree_lookup() but operates on an inode's in-memory
 354  * record list.
 355  *
 356  * The lookup must fail if the record is marked for deferred deletion.
 357  */
 358 static
 359 int
 360 hammer_mem_lookup(hammer_cursor_t cursor, hammer_inode_t ip)
 361 {
 362         int error;
 363
 364         if (cursor->iprec) {
 365                 hammer_rel_mem_record(cursor->iprec);
 366                 cursor->iprec = NULL;
 367         }
 368         if (cursor->ip) {
 369                 KKASSERT(cursor->ip->cursor_ip_refs > 0);
 370                 --cursor->ip->cursor_ip_refs;
 371 #if 0
 372                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 373                                                   &cursor->ip->rec_tree);
 374 #endif
 375         }
 376         cursor->ip = ip;
 377 #if 0
 378         hammer_rec_rb_tree_scan_info_link(&cursor->scan, &ip->rec_tree);
 379 #endif
 380         ++ip->cursor_ip_refs;
 381
 382 #if 0
 383         cursor->scan.node = NULL;
 384 #endif
 385         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_find_cmp,
 386                                    hammer_rec_scan_callback, cursor);
 387
 388         if (cursor->iprec == NULL)
 389                 error = ENOENT;
 390         else
 391                 error = 0;
 392         return(error);
 393 }
 394
 395 /*
 396  * hammer_mem_first() - locate the first in-memory record matching the
 397  * cursor within the bounds of the key range.
 398  */
 399 static
 400 int
 401 hammer_mem_first(hammer_cursor_t cursor, hammer_inode_t ip)
 402 {
 403         if (cursor->iprec) {
 404                 hammer_rel_mem_record(cursor->iprec);
 405                 cursor->iprec = NULL;
 406         }
 407         if (cursor->ip) {
 408                 KKASSERT(cursor->ip->cursor_ip_refs > 0);
 409                 --cursor->ip->cursor_ip_refs;
 410 #if 0
 411                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 412                                                   &cursor->ip->rec_tree);
 413 #endif
 414         }
 415         cursor->ip = ip;
 416 #if 0
 417         hammer_rec_rb_tree_scan_info_link(&cursor->scan, &ip->rec_tree);
 418 #endif
 419         ++ip->cursor_ip_refs;
 420
 421 #if 0
 422         cursor->scan.node = NULL;
 423 #endif
 424         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 425                                    hammer_rec_scan_callback, cursor);
 426
 427         /*
 428          * Adjust scan.node and keep it linked into the RB-tree so we can
 429          * hold the cursor through third party modifications of the RB-tree.
 430          */
 431         if (cursor->iprec) {
 432 #if 0
 433                 cursor->scan.node = hammer_rec_rb_tree_RB_NEXT(cursor->iprec);
 434 #endif
 435                 return(0);
 436         }
 437         return(ENOENT);
 438 }
 439
 440 void
 441 hammer_mem_done(hammer_cursor_t cursor)
 442 {
 443         if (cursor->ip) {
 444                 KKASSERT(cursor->ip->cursor_ip_refs > 0);
 445                 --cursor->ip->cursor_ip_refs;
 446 #if 0
 447                 hammer_rec_rb_tree_scan_info_done(&cursor->scan,
 448                                                   &cursor->ip->rec_tree);
 449 #endif
 450                 cursor->ip = NULL;
 451         }
 452         if (cursor->iprec) {
 453                 hammer_rel_mem_record(cursor->iprec);
 454                 cursor->iprec = NULL;
 455         }
 456 }
 457
 458 /************************************************************************
 459  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 460  ************************************************************************
 461  *
 462  * These functions manipulate in-memory records.  Such records typically
 463  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 464  */
 465
 466 /*
 467  * Add a directory entry (dip,ncp) which references inode (ip).
 468  *
 469  * Note that the low 32 bits of the namekey are set temporarily to create
 470  * a unique in-memory record, and may be modified a second time when the
 471  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 472  * all 0's when synching to disk, which is not handled here.
 473  */
 474 int
 475 hammer_ip_add_directory(struct hammer_transaction *trans,
 476                      struct hammer_inode *dip, struct namecache *ncp,
 477                      struct hammer_inode *ip)
 478 {
 479         hammer_record_t record;
 480         hammer_depend_t depend;
 481         int error;
 482         int bytes;
 483
 484         record = hammer_alloc_mem_record(dip);
 485         depend = kmalloc(sizeof(*depend), M_HAMMER, M_WAITOK|M_ZERO);
 486
 487         bytes = ncp->nc_nlen;   /* NOTE: terminating \0 is NOT included */
 488         if (++trans->hmp->namekey_iterator == 0)
 489                 ++trans->hmp->namekey_iterator;
 490
 491         record->rec.entry.base.base.obj_id = dip->obj_id;
 492         record->rec.entry.base.base.key =
 493                 hammer_directory_namekey(ncp->nc_name, bytes);
 494         record->rec.entry.base.base.key += trans->hmp->namekey_iterator;
 495         record->rec.entry.base.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 496         record->rec.entry.base.base.obj_type = ip->ino_rec.base.base.obj_type;
 497         record->rec.entry.obj_id = ip->obj_id;
 498         record->data = (void *)ncp->nc_name;
 499         record->rec.entry.base.data_len = bytes;
 500         ++ip->ino_rec.ino_nlinks;
 501         hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 502         /* NOTE: copies record->data */
 503
 504         /*
 505          * If the inode gets synced cause the directory entry
 506          * to be synced as well, or vise-versa.
 507          */
 508         hammer_ref(&record->lock);      /* for depend entry */
 509         hammer_ref(&ip->lock);          /* for depend entry */
 510         depend->ip = ip;
 511         depend->record = record;
 512         TAILQ_INSERT_TAIL(&ip->depend_list, depend, ip_entry);
 513         TAILQ_INSERT_TAIL(&record->depend_list, depend, rec_entry);
 514         ++ip->depend_count;
 515
 516         error = hammer_mem_add(trans, record);
 517         return(error);
 518 }
 519
 520 /*
 521  * Delete the directory entry and update the inode link count.  The
 522  * cursor must be seeked to the directory entry record being deleted.
 523  *
 524  * The related inode should be share-locked by the caller.  The caller is
 525  * on the frontend.
 526  *
 527  * This function can return EDEADLK requiring the caller to terminate
 528  * the cursor, any locks, wait on the returned record, and retry.
 529  */
 530 int
 531 hammer_ip_del_directory(struct hammer_transaction *trans,
 532                      hammer_cursor_t cursor, struct hammer_inode *dip,
 533                      struct hammer_inode *ip)
 534 {
 535         hammer_record_t record;
 536         hammer_depend_t depend;
 537         int error;
 538
 539         if (cursor->record == &cursor->iprec->rec) {
 540                 /*
 541                  * In-memory (unsynchronized) records can simply be freed.
 542                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 543                  * by the backend, we must still avoid races against the
 544                  * backend potentially syncing the record to the media.
 545                  *
 546                  * We cannot call hammer_ip_delete_record(), that routine may
 547                  * only be called from the backend.
 548                  */
 549                 record = cursor->iprec;
 550                 if (record->state == HAMMER_FST_FLUSH) {
 551                         KKASSERT(cursor->deadlk_rec == NULL);
 552                         hammer_ref(&record->lock);
 553                         cursor->deadlk_rec = record;
 554                         error = EDEADLK;
 555                 } else {
 556                         hammer_delete_mem_record(cursor->iprec);
 557                         error = 0;
 558                 }
 559         } else {
 560                 /*
 561                  * If the record is on-disk we have to queue the deletion by
 562                  * the record's key.  This also causes lookups to skip the
 563                  * record.
 564                  */
 565                 depend = kmalloc(sizeof(*depend), M_HAMMER, M_WAITOK|M_ZERO);
 566
 567                 record = hammer_alloc_mem_record(dip);
 568                 record->rec.entry.base.base = cursor->record->base.base;
 569                 hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 570                 record->flags |= HAMMER_RECF_DELETE_ONDISK;
 571
 572                 /*
 573                  * If the inode gets synced cause the directory entry
 574                  * to be synced as well, or vise-versa.
 575                  */
 576                 hammer_ref(&ip->lock);          /* for depend entry */
 577                 hammer_ref(&record->lock);      /* for depend entry */
 578                 depend->ip = ip;
 579                 depend->record = record;
 580                 TAILQ_INSERT_TAIL(&ip->depend_list, depend, ip_entry);
 581                 TAILQ_INSERT_TAIL(&record->depend_list, depend, rec_entry);
 582                 ++ip->depend_count;
 583
 584                 error = hammer_mem_add(trans, record);
 585         }
 586
 587         /*
 588          * One less link.  The file may still be open in the OS even after
 589          * all links have gone away so we only try to sync if the OS has
 590          * no references and nlinks falls to 0.
 591          *
 592          * We have to terminate the cursor before syncing the inode to
 593          * avoid deadlocking against ourselves.
 594          *
 595          * XXX we can't sync the inode here because the encompassing
 596          * transaction might be a rename and might update the inode
 597          * again with a new link.  That would force the delete_tid to be
 598          * the same as the create_tid and cause a panic.
 599          */
 600         if (error == 0) {
 601                 --ip->ino_rec.ino_nlinks;
 602                 hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 603                 if (ip->ino_rec.ino_nlinks == 0 &&
 604                     (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 605                         hammer_done_cursor(cursor);
 606                 }
 607
 608         }
 609         return(error);
 610 }
 611
 612 /*
 613  * Add a record to an inode.
 614  *
 615  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 616  * initialize the following additional fields:
 617  *
 618  * The related inode should be share-locked by the caller.  The caller is
 619  * on the frontend.
 620  *
 621  * record->rec.entry.base.base.key
 622  * record->rec.entry.base.base.rec_type
 623  * record->rec.entry.base.base.data_len
 624  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 625  */
 626 int
 627 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 628 {
 629         hammer_inode_t ip = record->ip;
 630         int error;
 631
 632         record->rec.base.base.obj_id = ip->obj_id;
 633         record->rec.base.base.obj_type = ip->ino_rec.base.base.obj_type;
 634
 635         hammer_modify_inode(trans, ip, HAMMER_INODE_RDIRTY);
 636         /* NOTE: copies record->data */
 637         error = hammer_mem_add(trans, record);
 638         return(error);
 639 }
 640
 641 /*
 642  * Sync data from a buffer cache buffer (typically) to the filesystem.  This
 643  * is called via the strategy called from a cached data source.  This code
 644  * is responsible for actually writing a data record out to the disk.
 645  *
 646  * This can only occur non-historically (i.e. 'current' data only).
 647  *
 648  * The file offset must be HAMMER_BUFSIZE aligned but the data length
 649  * can be truncated.  The record (currently) always represents a BUFSIZE
 650  * swath of space whether the data is truncated or not.
 651  */
 652 int
 653 hammer_ip_sync_data(hammer_transaction_t trans, hammer_inode_t ip,
 654                        int64_t offset, void *data, int bytes)
 655 {
 656         struct hammer_cursor cursor;
 657         hammer_record_ondisk_t rec;
 658         union hammer_btree_elm elm;
 659         hammer_off_t rec_offset;
 660         void *bdata;
 661         int error;
 662
 663         KKASSERT((offset & HAMMER_BUFMASK) == 0);
 664         KKASSERT(trans->type == HAMMER_TRANS_FLS);
 665 retry:
 666         error = hammer_init_cursor(trans, &cursor, &ip->cache[0]);
 667         if (error)
 668                 return(error);
 669         cursor.key_beg.obj_id = ip->obj_id;
 670         cursor.key_beg.key = offset + bytes;
 671         cursor.key_beg.create_tid = trans->tid;
 672         cursor.key_beg.delete_tid = 0;
 673         cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
 674         cursor.asof = trans->tid;
 675         cursor.flags |= HAMMER_CURSOR_INSERT;
 676
 677         /*
 678          * Issue a lookup to position the cursor.
 679          */
 680         error = hammer_btree_lookup(&cursor);
 681         if (error == 0) {
 682                 kprintf("hammer_ip_sync_data: duplicate data at "
 683                         "(%lld,%d) tid %016llx\n",
 684                         offset, bytes, trans->tid);
 685                 hammer_print_btree_elm(&cursor.node->ondisk->elms[cursor.index],
 686                                        HAMMER_BTREE_TYPE_LEAF, cursor.index);
 687                 panic("Duplicate data");
 688                 error = EIO;
 689         }
 690         if (error != ENOENT)
 691                 goto done;
 692
 693         /*
 694          * Allocate record and data space.  HAMMER_RECTYPE_DATA records
 695          * can cross buffer boundaries so we may have to split our bcopy.
 696          */
 697         rec = hammer_alloc_record(trans, &rec_offset, HAMMER_RECTYPE_DATA,
 698                                   &cursor.record_buffer,
 699                                   bytes, &bdata,
 700                                   &cursor.data_buffer, &error);
 701         if (rec == NULL)
 702                 goto done;
 703         if (hammer_debug_general & 0x1000)
 704                 kprintf("OOB RECOR2 DATA REC %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, rec->base.data_len);
 705
 706         /*
 707          * Fill everything in and insert our B-Tree node.
 708          *
 709          * NOTE: hammer_alloc_record() has already marked the related
 710          * buffers as modified.  If we do it again we will generate
 711          * unnecessary undo elements.
 712          */
 713         hammer_modify_buffer(trans, cursor.record_buffer, NULL, 0);
 714         rec->base.base.btype = HAMMER_BTREE_TYPE_RECORD;
 715         rec->base.base.obj_id = ip->obj_id;
 716         rec->base.base.key = offset + bytes;
 717         rec->base.base.create_tid = trans->tid;
 718         rec->base.base.delete_tid = 0;
 719         rec->base.base.rec_type = HAMMER_RECTYPE_DATA;
 720         rec->base.data_crc = crc32(data, bytes);
 721         hammer_modify_buffer_done(cursor.record_buffer);
 722         KKASSERT(rec->base.data_len == bytes);
 723
 724         hammer_modify_buffer(trans, cursor.data_buffer, NULL, 0);
 725         bcopy(data, bdata, bytes);
 726         hammer_modify_buffer_done(cursor.data_buffer);
 727
 728         elm.leaf.base = rec->base.base;
 729         elm.leaf.rec_offset = rec_offset;
 730         elm.leaf.data_offset = rec->base.data_off;
 731         elm.leaf.data_len = bytes;
 732         elm.leaf.data_crc = rec->base.data_crc;
 733
 734         /*
 735          * Data records can wind up on-disk before the inode itself is
 736          * on-disk.  One must assume data records may be on-disk if either
 737          * HAMMER_INODE_DONDISK or HAMMER_INODE_ONDISK is set
 738          */
 739         ip->flags |= HAMMER_INODE_DONDISK;
 740
 741         error = hammer_btree_insert(&cursor, &elm);
 742         if (error == 0)
 743                 goto done;
 744
 745         hammer_blockmap_free(trans, rec_offset, HAMMER_RECORD_SIZE);
 746 done:
 747         hammer_done_cursor(&cursor);
 748         if (error == EDEADLK)
 749                 goto retry;
 750         return(error);
 751 }
 752
 753 /*
 754  * Sync an in-memory record to the disk.  This is called by the backend.
 755  * This code is responsible for actually writing a record out to the disk.
 756  *
 757  * Any inode dependancies will queue the inode to the backend.
 758  *
 759  * NOTE: The frontend can mark the record deleted while it is queued to
 760  * the backend.  The deletion applies to a frontend operation and the
 761  * record must be treated as NOT having been deleted on the backend, so
 762  * we ignore the flag.
 763  */
 764 int
 765 hammer_ip_sync_record(hammer_transaction_t trans, hammer_record_t record)
 766 {
 767         struct hammer_cursor cursor;
 768         hammer_record_ondisk_t rec;
 769         union hammer_btree_elm elm;
 770         hammer_depend_t depend;
 771         hammer_off_t rec_offset;
 772         void *bdata;
 773         int error;
 774
 775         KKASSERT(record->state == HAMMER_FST_FLUSH);
 776
 777         while ((depend = TAILQ_FIRST(&record->depend_list)) != NULL) {
 778                 TAILQ_REMOVE(&record->depend_list, depend, rec_entry);
 779                 TAILQ_REMOVE(&depend->ip->depend_list, depend, ip_entry);
 780                 --depend->ip->depend_count;
 781                 kprintf("S");
 782                 hammer_flush_inode(depend->ip, 0);
 783                 hammer_rel_inode(depend->ip, 0);
 784                 hammer_unref(&record->lock);
 785                 KKASSERT(record->lock.refs > 0);
 786                 kfree(depend, M_HAMMER);
 787         }
 788
 789 retry:
 790         /*
 791          * Get a cursor, we will either be inserting or deleting.
 792          */
 793         error = hammer_init_cursor(trans, &cursor, &record->ip->cache[0]);
 794         if (error)
 795                 return(error);
 796         cursor.key_beg = record->rec.base.base;
 797
 798         /*
 799          * If we are deleting an exact match must be found on-disk.
 800          */
 801         if (record->flags & HAMMER_RECF_DELETE_ONDISK) {
 802                 error = hammer_btree_lookup(&cursor);
 803                 if (error == 0)
 804                         error = hammer_ip_delete_record(&cursor, trans->tid);
 805                 if (error == 0)
 806                         hammer_delete_mem_record(record);
 807                 goto done;
 808         }
 809
 810         /*
 811          * We are inserting.
 812          *
 813          * Issue a lookup to position the cursor and locate the cluster.  The
 814          * target key should not exist.  If we are creating a directory entry
 815          * we may have to iterate the low 32 bits of the key to find an unused
 816          * key.
 817          */
 818         cursor.flags |= HAMMER_CURSOR_INSERT;
 819
 820         for (;;) {
 821                 error = hammer_btree_lookup(&cursor);
 822                 if (error)
 823                         break;
 824                 if (record->rec.base.base.rec_type != HAMMER_RECTYPE_DIRENTRY) {
 825                         kprintf("hammer_ip_sync_record: duplicate rec "
 826                                 "at (%016llx)\n", record->rec.base.base.key);
 827                         Debugger("duplicate record1");
 828                         error = EIO;
 829                         break;
 830                 }
 831                 if (++trans->hmp->namekey_iterator == 0)
 832                         ++trans->hmp->namekey_iterator;
 833                 record->rec.base.base.key &= ~(0xFFFFFFFFLL);
 834                 record->rec.base.base.key |= trans->hmp->namekey_iterator;
 835                 cursor.key_beg.key = record->rec.base.base.key;
 836         }
 837         if (error != ENOENT)
 838                 goto done;
 839
 840         /*
 841          * Allocate the record and data.  The result buffers will be
 842          * marked as being modified and further calls to
 843          * hammer_modify_buffer() will result in unneeded UNDO records.
 844          *
 845          * Support zero-fill records (data == NULL and data_len != 0)
 846          */
 847         if (record->data == NULL) {
 848                 rec = hammer_alloc_record(trans, &rec_offset,
 849                                           record->rec.base.base.rec_type,
 850                                           &cursor.record_buffer,
 851                                           0, &bdata,
 852                                           NULL, &error);
 853                 if (hammer_debug_general & 0x1000)
 854                         kprintf("NULL RECORD DATA\n");
 855         } else if (record->flags & HAMMER_RECF_INBAND) {
 856                 rec = hammer_alloc_record(trans, &rec_offset,
 857                                           record->rec.base.base.rec_type,
 858                                           &cursor.record_buffer,
 859                                           record->rec.base.data_len, &bdata,
 860                                           NULL, &error);
 861                 if (hammer_debug_general & 0x1000)
 862                         kprintf("INBAND RECORD DATA %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, record->rec.base.data_len);
 863         } else {
 864                 rec = hammer_alloc_record(trans, &rec_offset,
 865                                           record->rec.base.base.rec_type,
 866                                           &cursor.record_buffer,
 867                                           record->rec.base.data_len, &bdata,
 868                                           &cursor.data_buffer, &error);
 869                 if (hammer_debug_general & 0x1000)
 870                         kprintf("OOB RECORD DATA REC %016llx DATA %016llx LEN=%d\n", rec_offset, rec->base.data_off, record->rec.base.data_len);
 871         }
 872
 873         if (rec == NULL)
 874                 goto done;
 875
 876         /*
 877          * Fill in the remaining fields and insert our B-Tree node.
 878          */
 879         hammer_modify_buffer(trans, cursor.record_buffer, NULL, 0);
 880         rec->base.base = record->rec.base.base;
 881         bcopy(&record->rec.base + 1, &rec->base + 1,
 882               HAMMER_RECORD_SIZE - sizeof(record->rec.base));
 883
 884         /*
 885          * Copy the data and deal with zero-fill support.
 886          */
 887         if (record->data && (record->flags & HAMMER_RECF_INBAND)) {
 888                 rec->base.data_crc = crc32(record->data, rec->base.data_len);
 889                 bcopy(record->data, bdata, rec->base.data_len);
 890         } else if (record->data) {
 891                 rec->base.data_crc = crc32(record->data, rec->base.data_len);
 892                 hammer_modify_buffer(trans, cursor.data_buffer, NULL, 0);
 893                 bcopy(record->data, bdata, rec->base.data_len);
 894                 hammer_modify_buffer_done(cursor.data_buffer);
 895         } else {
 896                 rec->base.data_len = record->rec.base.data_len;
 897         }
 898         hammer_modify_buffer_done(cursor.record_buffer);
 899
 900         elm.leaf.base = record->rec.base.base;
 901         elm.leaf.rec_offset = rec_offset;
 902         elm.leaf.data_offset = rec->base.data_off;
 903         elm.leaf.data_len = rec->base.data_len;
 904         elm.leaf.data_crc = rec->base.data_crc;
 905
 906         error = hammer_btree_insert(&cursor, &elm);
 907
 908         /*
 909          * Clean up on success, or fall through on error.
 910          */
 911         if (error == 0) {
 912                 hammer_delete_mem_record(record);
 913                 goto done;
 914         }
 915
 916         /*
 917          * Try to unwind the allocation
 918          */
 919         hammer_blockmap_free(trans, rec_offset, HAMMER_RECORD_SIZE);
 920 done:
 921         hammer_done_cursor(&cursor);
 922         if (error == EDEADLK)
 923                 goto retry;
 924         return(error);
 925 }
 926
 927 /*
 928  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
 929  * entry's key is used to deal with hash collisions in the upper 32 bits.
 930  * A unique 64 bit key is generated in-memory and may be regenerated a
 931  * second time when the directory record is flushed to the on-disk B-Tree.
 932  *
 933  * A referenced record is passed to this function.  This function
 934  * eats the reference.  If an error occurs the record will be deleted.
 935  *
 936  * A copy of the temporary record->data pointer provided by the caller
 937  * will be made.
 938  */
 939 static
 940 int
 941 hammer_mem_add(struct hammer_transaction *trans, hammer_record_t record)
 942 {
 943         void *data;
 944         int bytes;
 945         int reclen;
 946
 947         /*
 948          * Make a private copy of record->data
 949          */
 950         if (record->data) {
 951                 /*
 952                  * Try to embed the data in extra space in the record
 953                  * union, otherwise allocate a copy.
 954                  */
 955                 bytes = record->rec.base.data_len;
 956                 switch(record->rec.base.base.rec_type) {
 957                 case HAMMER_RECTYPE_DIRENTRY:
 958                         reclen = offsetof(struct hammer_entry_record, name[0]);
 959                         break;
 960                 case HAMMER_RECTYPE_DATA:
 961                         reclen = offsetof(struct hammer_data_record, data[0]);
 962                         break;
 963                 default:
 964                         reclen = sizeof(record->rec);
 965                         break;
 966                 }
 967                 if (reclen + bytes <= HAMMER_RECORD_SIZE) {
 968                         bcopy(record->data, (char *)&record->rec + reclen,
 969                               bytes);
 970                         record->data = (void *)((char *)&record->rec + reclen);
 971                         record->flags |= HAMMER_RECF_INBAND;
 972                 } else {
 973                         ++hammer_count_record_datas;
 974                         data = kmalloc(bytes, M_HAMMER, M_WAITOK);
 975                         record->flags |= HAMMER_RECF_ALLOCDATA;
 976                         bcopy(record->data, data, bytes);
 977                         record->data = data;
 978                 }
 979         }
 980
 981         /*
 982          * Insert into the RB tree, find an unused iterator if this is
 983          * a directory entry.
 984          */
 985         while (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
 986                 if (record->rec.base.base.rec_type != HAMMER_RECTYPE_DIRENTRY){
 987                         hammer_delete_mem_record(record);
 988                         hammer_rel_mem_record(record);
 989                         return (EEXIST);
 990                 }
 991                 if (++trans->hmp->namekey_iterator == 0)
 992                         ++trans->hmp->namekey_iterator;
 993                 record->rec.base.base.key &= ~(0xFFFFFFFFLL);
 994                 record->rec.base.base.key |= trans->hmp->namekey_iterator;
 995         }
 996         record->flags |= HAMMER_RECF_ONRBTREE;
 997         hammer_modify_inode(trans, record->ip, HAMMER_INODE_XDIRTY);
 998         hammer_rel_mem_record(record);
 999         return(0);
1000 }
1001
1002 /************************************************************************
1003  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1004  ************************************************************************
1005  *
1006  * These functions augment the B-Tree scanning functions in hammer_btree.c
1007  * by merging in-memory records with on-disk records.
1008  */
1009
1010 /*
1011  * Locate a particular record either in-memory or on-disk.
1012  *
1013  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1014  * NOT be called to iterate results.
1015  */
1016 int
1017 hammer_ip_lookup(hammer_cursor_t cursor, struct hammer_inode *ip)
1018 {
1019         int error;
1020
1021         /*
1022          * If the element is in-memory return it without searching the
1023          * on-disk B-Tree
1024          */
1025         error = hammer_mem_lookup(cursor, ip);
1026         if (error == 0) {
1027                 cursor->record = &cursor->iprec->rec;
1028                 return(error);
1029         }
1030         if (error != ENOENT)
1031                 return(error);
1032
1033         /*
1034          * If the inode has on-disk components search the on-disk B-Tree.
1035          */
1036         if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1037                 return(error);
1038         error = hammer_btree_lookup(cursor);
1039         if (error == 0)
1040                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
1041         return(error);
1042 }
1043
1044 /*
1045  * Locate the first record within the cursor's key_beg/key_end range,
1046  * restricted to a particular inode.  0 is returned on success, ENOENT
1047  * if no records matched the requested range, or some other error.
1048  *
1049  * When 0 is returned hammer_ip_next() may be used to iterate additional
1050  * records within the requested range.
1051  *
1052  * This function can return EDEADLK, requiring the caller to terminate
1053  * the cursor and try again.
1054  */
1055 int
1056 hammer_ip_first(hammer_cursor_t cursor, struct hammer_inode *ip)
1057 {
1058         int error;
1059
1060         /*
1061          * Clean up fields and setup for merged scan
1062          */
1063         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1064         cursor->flags |= HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM;
1065         cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_MEMEOF;
1066         if (cursor->iprec) {
1067                 hammer_rel_mem_record(cursor->iprec);
1068                 cursor->iprec = NULL;
1069         }
1070
1071         /*
1072          * Search the on-disk B-Tree.  hammer_btree_lookup() only does an
1073          * exact lookup so if we get ENOENT we have to call the iterate
1074          * function to validate the first record after the begin key.
1075          *
1076          * The ATEDISK flag is used by hammer_btree_iterate to determine
1077          * whether it must index forwards or not.  It is also used here
1078          * to select the next record from in-memory or on-disk.
1079          *
1080          * EDEADLK can only occur if the lookup hit an empty internal
1081          * element and couldn't delete it.  Since this could only occur
1082          * in-range, we can just iterate from the failure point.
1083          */
1084         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1085                 error = hammer_btree_lookup(cursor);
1086                 if (error == ENOENT || error == EDEADLK) {
1087                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1088                         if (hammer_debug_general & 0x2000)
1089                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1090                         error = hammer_btree_iterate(cursor);
1091                 }
1092                 if (error && error != ENOENT)
1093                         return(error);
1094                 if (error == 0) {
1095                         cursor->flags &= ~HAMMER_CURSOR_DISKEOF;
1096                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1097                 } else {
1098                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1099                 }
1100         }
1101
1102         /*
1103          * Search the in-memory record list (Red-Black tree).  Unlike the
1104          * B-Tree search, mem_first checks for records in the range.
1105          */
1106         error = hammer_mem_first(cursor, ip);
1107         if (error && error != ENOENT)
1108                 return(error);
1109         if (error == 0) {
1110                 cursor->flags &= ~HAMMER_CURSOR_MEMEOF;
1111                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1112                 if (hammer_ip_iterate_mem_good(cursor, cursor->iprec) == 0)
1113                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1114         }
1115
1116         /*
1117          * This will return the first matching record.
1118          */
1119         return(hammer_ip_next(cursor));
1120 }
1121
1122 /*
1123  * Retrieve the next record in a merged iteration within the bounds of the
1124  * cursor.  This call may be made multiple times after the cursor has been
1125  * initially searched with hammer_ip_first().
1126  *
1127  * 0 is returned on success, ENOENT if no further records match the
1128  * requested range, or some other error code is returned.
1129  */
1130 int
1131 hammer_ip_next(hammer_cursor_t cursor)
1132 {
1133         hammer_btree_elm_t elm;
1134         hammer_record_t rec, save;
1135         int error;
1136         int r;
1137
1138 next_btree:
1139         /*
1140          * Load the current on-disk and in-memory record.  If we ate any
1141          * records we have to get the next one.
1142          *
1143          * If we deleted the last on-disk record we had scanned ATEDISK will
1144          * be clear and DELBTREE will be set, forcing a call to iterate. The
1145          * fact that ATEDISK is clear causes iterate to re-test the 'current'
1146          * element.  If ATEDISK is set, iterate will skip the 'current'
1147          * element.
1148          *
1149          * Get the next on-disk record
1150          */
1151         if (cursor->flags & (HAMMER_CURSOR_ATEDISK|HAMMER_CURSOR_DELBTREE)) {
1152                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1153                         error = hammer_btree_iterate(cursor);
1154                         cursor->flags &= ~HAMMER_CURSOR_DELBTREE;
1155                         if (error == 0)
1156                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1157                         else
1158                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1159                                                  HAMMER_CURSOR_ATEDISK;
1160                 }
1161         }
1162
1163 next_memory:
1164         /*
1165          * Get the next in-memory record.  The record can be ripped out
1166          * of the RB tree so we maintain a scan_info structure to track
1167          * the next node.
1168          *
1169          * hammer_rec_scan_cmp:  Is the record still in our general range,
1170          *                       (non-inclusive of snapshot exclusions)?
1171          * hammer_rec_scan_callback: Is the record in our snapshot?
1172          */
1173         if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1174                 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1175                         save = cursor->iprec;
1176                         cursor->iprec = NULL;
1177                         rec = save ? hammer_rec_rb_tree_RB_NEXT(save) : NULL;
1178                         while (rec) {
1179                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1180                                         break;
1181                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1182                                         break;
1183                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1184                         }
1185                         if (save)
1186                                 hammer_rel_mem_record(save);
1187                         if (cursor->iprec) {
1188                                 KKASSERT(cursor->iprec == rec);
1189                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1190 #if 0
1191                                 cursor->scan.node =
1192                                         hammer_rec_rb_tree_RB_NEXT(rec);
1193 #endif
1194                         } else {
1195                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1196                         }
1197                 }
1198         }
1199
1200         /*
1201          * Extract either the disk or memory record depending on their
1202          * relative position.
1203          */
1204         error = 0;
1205         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1206         case 0:
1207                 /*
1208                  * Both entries valid
1209                  */
1210                 elm = &cursor->node->ondisk->elms[cursor->index];
1211                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->rec.base.base);
1212                 if (r < 0) {
1213                         error = hammer_btree_extract(cursor,
1214                                                      HAMMER_CURSOR_GET_RECORD);
1215                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1216                         break;
1217                 }
1218
1219                 /*
1220                  * If the entries match the memory entry must specify
1221                  * an on-disk deletion.  Eat both entries unless the
1222                  * caller wants visibility into the special records.
1223                  */
1224                 if (r == 0) {
1225                         KKASSERT(cursor->iprec->flags &
1226                                  HAMMER_RECF_DELETE_ONDISK);
1227                         if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1228                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1229                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1230                                 goto next_btree;
1231                         }
1232                 }
1233                 /* fall through to the memory entry */
1234         case HAMMER_CURSOR_ATEDISK:
1235                 /*
1236                  * Only the memory entry is valid.  If the record is
1237                  * placemarking an on-disk deletion, we skip it unless
1238                  * the caller wants special record visibility.
1239                  */
1240                 cursor->record = &cursor->iprec->rec;
1241                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1242                 if (cursor->iprec->flags & HAMMER_RECF_DELETE_ONDISK) {
1243                         if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0)
1244                                 goto next_memory;
1245                 }
1246                 break;
1247         case HAMMER_CURSOR_ATEMEM:
1248                 /*
1249                  * Only the disk entry is valid
1250                  */
1251                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
1252                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1253                 break;
1254         default:
1255                 /*
1256                  * Neither entry is valid
1257                  *
1258                  * XXX error not set properly
1259                  */
1260                 cursor->record = NULL;
1261                 error = ENOENT;
1262                 break;
1263         }
1264         return(error);
1265 }
1266
1267 /*
1268  * Resolve the cursor->data pointer for the current cursor position in
1269  * a merged iteration.
1270  */
1271 int
1272 hammer_ip_resolve_data(hammer_cursor_t cursor)
1273 {
1274         int error;
1275
1276         if (cursor->iprec && cursor->record == &cursor->iprec->rec) {
1277                 cursor->data = cursor->iprec->data;
1278                 error = 0;
1279         } else {
1280                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1281         }
1282         return(error);
1283 }
1284
1285 int
1286 hammer_ip_resolve_record_and_data(hammer_cursor_t cursor)
1287 {
1288         int error;
1289
1290         if (cursor->iprec && cursor->record == &cursor->iprec->rec) {
1291                 cursor->data = cursor->iprec->data;
1292                 error = 0;
1293         } else {
1294                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA |
1295                                                      HAMMER_CURSOR_GET_RECORD);
1296         }
1297         return(error);
1298 }
1299
1300 /*
1301  * Delete all records within the specified range for inode ip.
1302  *
1303  * NOTE: An unaligned range will cause new records to be added to cover
1304  * the edge cases. (XXX not implemented yet).
1305  *
1306  * NOTE: ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1307  *
1308  * NOTE: Record keys for regular file data have to be special-cased since
1309  * they indicate the end of the range (key = base + bytes).
1310  */
1311 int
1312 hammer_ip_delete_range(hammer_transaction_t trans, hammer_inode_t ip,
1313                        int64_t ran_beg, int64_t ran_end)
1314 {
1315         struct hammer_cursor cursor;
1316         hammer_record_ondisk_t rec;
1317         hammer_base_elm_t base;
1318         int error;
1319         int64_t off;
1320
1321 #if 0
1322         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1323 #endif
1324
1325         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1326 retry:
1327         hammer_init_cursor(trans, &cursor, &ip->cache[0]);
1328
1329         cursor.key_beg.obj_id = ip->obj_id;
1330         cursor.key_beg.create_tid = 0;
1331         cursor.key_beg.delete_tid = 0;
1332         cursor.key_beg.obj_type = 0;
1333         cursor.asof = ip->obj_asof;
1334         cursor.flags |= HAMMER_CURSOR_ASOF;
1335         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1336         cursor.flags |= HAMMER_CURSOR_BACKEND;
1337
1338         cursor.key_end = cursor.key_beg;
1339         if (ip->ino_rec.base.base.obj_type == HAMMER_OBJTYPE_DBFILE) {
1340                 cursor.key_beg.key = ran_beg;
1341                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
1342                 cursor.key_end.rec_type = HAMMER_RECTYPE_DB;
1343                 cursor.key_end.key = ran_end;
1344         } else {
1345                 /*
1346                  * The key in the B-Tree is (base+bytes), so the first possible
1347                  * matching key is ran_beg + 1.
1348                  */
1349                 int64_t tmp64;
1350
1351                 cursor.key_beg.key = ran_beg + 1;
1352                 cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
1353                 cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
1354
1355                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1356                 if (tmp64 < ran_end)
1357                         cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1358                 else
1359                         cursor.key_end.key = ran_end + MAXPHYS + 1;
1360         }
1361         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
1362
1363         error = hammer_ip_first(&cursor, ip);
1364
1365         /*
1366          * Iterate through matching records and mark them as deleted.
1367          */
1368         while (error == 0) {
1369                 rec = cursor.record;
1370                 base = &rec->base.base;
1371
1372                 KKASSERT(base->delete_tid == 0);
1373
1374                 /*
1375                  * There may be overlap cases for regular file data.  Also
1376                  * remember the key for a regular file record is the offset
1377                  * of the last byte of the record (base + len - 1), NOT the
1378                  * base offset.
1379                  */
1380 #if 0
1381                 kprintf("delete_range rec_type %02x\n", base->rec_type);
1382 #endif
1383                 if (base->rec_type == HAMMER_RECTYPE_DATA) {
1384 #if 0
1385                         kprintf("delete_range loop key %016llx,%d\n",
1386                                 base->key - rec->base.data_len, rec->base.data_len);
1387 #endif
1388                         off = base->key - rec->base.data_len;
1389                         /*
1390                          * Check the left edge case.  We currently do not
1391                          * split existing records.
1392                          */
1393                         if (off < ran_beg) {
1394                                 panic("hammer left edge case %016llx %d\n",
1395                                         base->key, rec->base.data_len);
1396                         }
1397
1398                         /*
1399                          * Check the right edge case.  Note that the
1400                          * record can be completely out of bounds, which
1401                          * terminates the search.
1402                          *
1403                          * base->key is exclusive of the right edge while
1404                          * ran_end is inclusive of the right edge.  The
1405                          * (key - data_len) left boundary is inclusive.
1406                          *
1407                          * XXX theory-check this test at some point, are
1408                          * we missing a + 1 somewhere?  Note that ran_end
1409                          * could overflow.
1410                          */
1411                         if (base->key - 1 > ran_end) {
1412                                 if (base->key - rec->base.data_len > ran_end)
1413                                         break;
1414                                 panic("hammer right edge case\n");
1415                         }
1416                 }
1417
1418                 /*
1419                  * Mark the record and B-Tree entry as deleted.  This will
1420                  * also physically delete the B-Tree entry, record, and
1421                  * data if the retention policy dictates.  The function
1422                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1423                  * uses to perform a fixup.
1424                  */
1425                 error = hammer_ip_delete_record(&cursor, trans->tid);
1426                 if (error)
1427                         break;
1428                 error = hammer_ip_next(&cursor);
1429         }
1430         hammer_done_cursor(&cursor);
1431         if (error == EDEADLK)
1432                 goto retry;
1433         if (error == ENOENT)
1434                 error = 0;
1435         return(error);
1436 }
1437
1438 /*
1439  * Delete all records associated with an inode except the inode record
1440  * itself.
1441  */
1442 int
1443 hammer_ip_delete_range_all(hammer_transaction_t trans, hammer_inode_t ip)
1444 {
1445         struct hammer_cursor cursor;
1446         hammer_record_ondisk_t rec;
1447         hammer_base_elm_t base;
1448         int error;
1449
1450         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1451 retry:
1452         hammer_init_cursor(trans, &cursor, &ip->cache[0]);
1453
1454         cursor.key_beg.obj_id = ip->obj_id;
1455         cursor.key_beg.create_tid = 0;
1456         cursor.key_beg.delete_tid = 0;
1457         cursor.key_beg.obj_type = 0;
1458         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1459         cursor.key_beg.key = HAMMER_MIN_KEY;
1460
1461         cursor.key_end = cursor.key_beg;
1462         cursor.key_end.rec_type = 0xFFFF;
1463         cursor.key_end.key = HAMMER_MAX_KEY;
1464
1465         cursor.asof = ip->obj_asof;
1466         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1467         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1468         cursor.flags |= HAMMER_CURSOR_BACKEND;
1469
1470         error = hammer_ip_first(&cursor, ip);
1471
1472         /*
1473          * Iterate through matching records and mark them as deleted.
1474          */
1475         while (error == 0) {
1476                 rec = cursor.record;
1477                 base = &rec->base.base;
1478
1479                 KKASSERT(base->delete_tid == 0);
1480
1481                 /*
1482                  * Mark the record and B-Tree entry as deleted.  This will
1483                  * also physically delete the B-Tree entry, record, and
1484                  * data if the retention policy dictates.  The function
1485                  * will set HAMMER_CURSOR_DELBTREE which hammer_ip_next()
1486                  * uses to perform a fixup.
1487                  */
1488                 error = hammer_ip_delete_record(&cursor, trans->tid);
1489                 if (error)
1490                         break;
1491                 error = hammer_ip_next(&cursor);
1492         }
1493         hammer_done_cursor(&cursor);
1494         if (error == EDEADLK)
1495                 goto retry;
1496         if (error == ENOENT)
1497                 error = 0;
1498         return(error);
1499 }
1500
1501 /*
1502  * Delete the record at the current cursor.  On success the cursor will
1503  * be positioned appropriately for an iteration but may no longer be at
1504  * a leaf node.
1505  *
1506  * This routine is only called from the backend.
1507  *
1508  * NOTE: This can return EDEADLK, requiring the caller to terminate the
1509  * cursor and retry.
1510  */
1511 int
1512 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_tid_t tid)
1513 {
1514         hammer_btree_elm_t elm;
1515         hammer_mount_t hmp;
1516         int error;
1517         int dodelete;
1518
1519         /*
1520          * In-memory (unsynchronized) records can simply be freed.
1521          */
1522         if (cursor->record == &cursor->iprec->rec) {
1523                 hammer_delete_mem_record(cursor->iprec);
1524                 cursor->iprec->flags |= HAMMER_RECF_DELETED_BE;
1525                 return(0);
1526         }
1527
1528         /*
1529          * On-disk records are marked as deleted by updating their delete_tid.
1530          * This does not effect their position in the B-Tree (which is based
1531          * on their create_tid).
1532          */
1533         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_RECORD);
1534         elm = NULL;
1535         hmp = cursor->node->hmp;
1536
1537         dodelete = 0;
1538         if (error == 0) {
1539                 error = hammer_cursor_upgrade(cursor);
1540                 if (error == 0) {
1541                         elm = &cursor->node->ondisk->elms[cursor->index];
1542                         hammer_modify_node(cursor->trans, cursor->node,
1543                                            &elm->leaf.base.delete_tid,
1544                                            sizeof(elm->leaf.base.delete_tid));
1545                         elm->leaf.base.delete_tid = tid;
1546                         hammer_modify_node_done(cursor->node);
1547
1548                         /*
1549                          * An on-disk record cannot have the same delete_tid
1550                          * as its create_tid.  In a chain of record updates
1551                          * this could result in a duplicate record.
1552                          */
1553                         KKASSERT(elm->leaf.base.delete_tid != elm->leaf.base.create_tid);
1554                         hammer_modify_buffer(cursor->trans, cursor->record_buffer, &cursor->record->base.base.delete_tid, sizeof(hammer_tid_t));
1555                         cursor->record->base.base.delete_tid = tid;
1556                         hammer_modify_buffer_done(cursor->record_buffer);
1557                 }
1558         }
1559
1560         /*
1561          * If we were mounted with the nohistory option, we physically
1562          * delete the record.
1563          */
1564         if (hmp->hflags & HMNT_NOHISTORY)
1565                 dodelete = 1;
1566
1567         if (error == 0 && dodelete) {
1568                 error = hammer_delete_at_cursor(cursor, NULL);
1569                 if (error) {
1570                         panic("hammer_ip_delete_record: unable to physically delete the record!\n");
1571                         error = 0;
1572                 }
1573         }
1574         return(error);
1575 }
1576
1577 int
1578 hammer_delete_at_cursor(hammer_cursor_t cursor, int64_t *stat_bytes)
1579 {
1580         hammer_btree_elm_t elm;
1581         hammer_off_t rec_offset;
1582         hammer_off_t data_offset;
1583         int32_t data_len;
1584         u_int16_t rec_type;
1585         int error;
1586
1587         elm = &cursor->node->ondisk->elms[cursor->index];
1588         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
1589
1590         rec_offset = elm->leaf.rec_offset;
1591         data_offset = elm->leaf.data_offset;
1592         data_len = elm->leaf.data_len;
1593         rec_type = elm->leaf.base.rec_type;
1594
1595         error = hammer_btree_delete(cursor);
1596         if (error == 0) {
1597                 /*
1598                  * This forces a fixup for the iteration because
1599                  * the cursor is now either sitting at the 'next'
1600                  * element or sitting at the end of a leaf.
1601                  */
1602                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1603                         cursor->flags |= HAMMER_CURSOR_DELBTREE;
1604                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1605                 }
1606         }
1607         if (error == 0) {
1608                 hammer_blockmap_free(cursor->trans, rec_offset,
1609                                      sizeof(union hammer_record_ondisk));
1610         }
1611         if (error == 0) {
1612                 switch(data_offset & HAMMER_OFF_ZONE_MASK) {
1613                 case HAMMER_ZONE_LARGE_DATA:
1614                 case HAMMER_ZONE_SMALL_DATA:
1615                         hammer_blockmap_free(cursor->trans,
1616                                              data_offset, data_len);
1617                         break;
1618                 default:
1619                         break;
1620                 }
1621         }
1622 #if 0
1623         kprintf("hammer_delete_at_cursor: %d:%d:%08x %08x/%d "
1624                 "(%d remain in cluster)\n",
1625                 cluster->volume->vol_no, cluster->clu_no,
1626                 rec_offset, data_offset, data_len,
1627                 cluster->ondisk->stat_records);
1628 #endif
1629         return (error);
1630 }
1631
1632 /*
1633  * Determine whether a directory is empty or not.  Returns 0 if the directory
1634  * is empty, ENOTEMPTY if it isn't, plus other possible errors.
1635  */
1636 int
1637 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
1638 {
1639         struct hammer_cursor cursor;
1640         int error;
1641
1642         hammer_init_cursor(trans, &cursor, &ip->cache[0]);
1643
1644         cursor.key_beg.obj_id = ip->obj_id;
1645         cursor.key_beg.create_tid = 0;
1646         cursor.key_beg.delete_tid = 0;
1647         cursor.key_beg.obj_type = 0;
1648         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
1649         cursor.key_beg.key = HAMMER_MIN_KEY;
1650
1651         cursor.key_end = cursor.key_beg;
1652         cursor.key_end.rec_type = 0xFFFF;
1653         cursor.key_end.key = HAMMER_MAX_KEY;
1654
1655         cursor.asof = ip->obj_asof;
1656         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
1657
1658         error = hammer_ip_first(&cursor, ip);
1659         if (error == ENOENT)
1660                 error = 0;
1661         else if (error == 0)
1662                 error = ENOTEMPTY;
1663         hammer_done_cursor(&cursor);
1664         return(error);
1665 }
1666