sys/vfs/hammer/hammer_object.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_object.c,v 1.97 2008/09/23 22:28:56 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_lookup(hammer_cursor_t cursor);
  40 static void hammer_mem_first(hammer_cursor_t cursor);
  41 static int hammer_frontend_trunc_callback(hammer_record_t record,
  42                                 void *data __unused);
  43 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
  46                       hammer_btree_leaf_elm_t leaf);
  47
  48 struct rec_trunc_info {
  49         u_int16_t       rec_type;
  50         int64_t         trunc_off;
  51 };
  52
  53 struct hammer_bulk_info {
  54         hammer_record_t record;
  55         struct hammer_btree_leaf_elm leaf;
  56 };
  57
  58 /*
  59  * Red-black tree support.  Comparison code for insertion.
  60  */
  61 static int
  62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  63 {
  64         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  65                 return(-1);
  66         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  67                 return(1);
  68
  69         if (rec1->leaf.base.key < rec2->leaf.base.key)
  70                 return(-1);
  71         if (rec1->leaf.base.key > rec2->leaf.base.key)
  72                 return(1);
  73
  74         /*
  75          * For search & insertion purposes records deleted by the
  76          * frontend or deleted/committed by the backend are silently
  77          * ignored.  Otherwise pipelined insertions will get messed
  78          * up.
  79          *
  80          * rec1 is greater then rec2 if rec1 is marked deleted.
  81          * rec1 is less then rec2 if rec2 is marked deleted.
  82          *
  83          * Multiple deleted records may be present, do not return 0
  84          * if both are marked deleted.
  85          */
  86         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  87                            HAMMER_RECF_COMMITTED)) {
  88                 return(1);
  89         }
  90         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  91                            HAMMER_RECF_COMMITTED)) {
  92                 return(-1);
  93         }
  94
  95         return(0);
  96 }
  97
  98 /*
  99  * Basic record comparison code similar to hammer_btree_cmp().
 100  */
 101 static int
 102 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 103 {
 104         if (elm->rec_type < rec->leaf.base.rec_type)
 105                 return(-3);
 106         if (elm->rec_type > rec->leaf.base.rec_type)
 107                 return(3);
 108
 109         if (elm->key < rec->leaf.base.key)
 110                 return(-2);
 111         if (elm->key > rec->leaf.base.key)
 112                 return(2);
 113
 114         /*
 115          * Never match against an item deleted by the frontend
 116          * or backend, or committed by the backend.
 117          *
 118          * elm is less then rec if rec is marked deleted.
 119          */
 120         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 121                           HAMMER_RECF_COMMITTED)) {
 122                 return(-1);
 123         }
 124         return(0);
 125 }
 126
 127 /*
 128  * Ranged scan to locate overlapping record(s).  This is used by
 129  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 130  * to use a ranged scan because the keys for data records with the
 131  * same file base offset can be different due to differing data_len's.
 132  *
 133  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 134  */
 135 static int
 136 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 137 {
 138         struct hammer_bulk_info *info = data;
 139         hammer_btree_leaf_elm_t leaf = &info->leaf;
 140
 141         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 142                 return(-3);
 143         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 144                 return(3);
 145
 146         /*
 147          * Overlap compare
 148          */
 149         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 150                 /* rec_beg >= leaf_end */
 151                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 152                         return(2);
 153                 /* rec_end <= leaf_beg */
 154                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 155                         return(-2);
 156         } else {
 157                 if (rec->leaf.base.key < leaf->base.key)
 158                         return(-2);
 159                 if (rec->leaf.base.key > leaf->base.key)
 160                         return(2);
 161         }
 162
 163         /*
 164          * We have to return 0 at this point, even if DELETED_FE is set,
 165          * because returning anything else will cause the scan to ignore
 166          * one of the branches when we really want it to check both.
 167          */
 168         return(0);
 169 }
 170
 171 /*
 172  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 173  * is reversed so the comparison result has to be negated.  key_beg and
 174  * key_end are both range-inclusive.
 175  *
 176  * Localized deletions are not cached in-memory.
 177  */
 178 static
 179 int
 180 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 181 {
 182         hammer_cursor_t cursor = data;
 183         int r;
 184
 185         r = hammer_rec_cmp(&cursor->key_beg, rec);
 186         if (r > 1)
 187                 return(-1);
 188         r = hammer_rec_cmp(&cursor->key_end, rec);
 189         if (r < -1)
 190                 return(1);
 191         return(0);
 192 }
 193
 194 /*
 195  * This compare function is used when simply looking up key_beg.
 196  */
 197 static
 198 int
 199 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 200 {
 201         hammer_cursor_t cursor = data;
 202         int r;
 203
 204         r = hammer_rec_cmp(&cursor->key_beg, rec);
 205         if (r > 1)
 206                 return(-1);
 207         if (r < -1)
 208                 return(1);
 209         return(0);
 210 }
 211
 212 /*
 213  * Locate blocks within the truncation range.  Partial blocks do not count.
 214  */
 215 static
 216 int
 217 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 218 {
 219         struct rec_trunc_info *info = data;
 220
 221         if (rec->leaf.base.rec_type < info->rec_type)
 222                 return(-1);
 223         if (rec->leaf.base.rec_type > info->rec_type)
 224                 return(1);
 225
 226         switch(rec->leaf.base.rec_type) {
 227         case HAMMER_RECTYPE_DB:
 228                 /*
 229                  * DB record key is not beyond the truncation point, retain.
 230                  */
 231                 if (rec->leaf.base.key < info->trunc_off)
 232                         return(-1);
 233                 break;
 234         case HAMMER_RECTYPE_DATA:
 235                 /*
 236                  * DATA record offset start is not beyond the truncation point,
 237                  * retain.
 238                  */
 239                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 240                         return(-1);
 241                 break;
 242         default:
 243                 panic("hammer_rec_trunc_cmp: unexpected record type");
 244         }
 245
 246         /*
 247          * The record start is >= the truncation point, return match,
 248          * the record should be destroyed.
 249          */
 250         return(0);
 251 }
 252
 253 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 254
 255 /*
 256  * Allocate a record for the caller to finish filling in.  The record is
 257  * returned referenced.
 258  */
 259 hammer_record_t
 260 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 261 {
 262         hammer_record_t record;
 263         hammer_mount_t hmp;
 264
 265         hmp = ip->hmp;
 266         ++hammer_count_records;
 267         record = kmalloc(sizeof(*record), hmp->m_misc,
 268                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 269         record->flush_state = HAMMER_FST_IDLE;
 270         record->ip = ip;
 271         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 272         record->leaf.data_len = data_len;
 273         hammer_ref(&record->lock);
 274
 275         if (data_len) {
 276                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
 277                 record->flags |= HAMMER_RECF_ALLOCDATA;
 278                 ++hammer_count_record_datas;
 279         }
 280
 281         return (record);
 282 }
 283
 284 void
 285 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 286 {
 287         while (record->flush_state == HAMMER_FST_FLUSH) {
 288                 record->flags |= HAMMER_RECF_WANTED;
 289                 tsleep(record, 0, ident, 0);
 290         }
 291 }
 292
 293 /*
 294  * Called from the backend, hammer_inode.c, after a record has been
 295  * flushed to disk.  The record has been exclusively locked by the
 296  * caller and interlocked with BE.
 297  *
 298  * We clean up the state, unlock, and release the record (the record
 299  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 300  */
 301 void
 302 hammer_flush_record_done(hammer_record_t record, int error)
 303 {
 304         hammer_inode_t target_ip;
 305
 306         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 307         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 308
 309         /*
 310          * If an error occured, the backend was unable to sync the
 311          * record to its media.  Leave the record intact.
 312          */
 313         if (error) {
 314                 hammer_critical_error(record->ip->hmp, record->ip, error,
 315                                       "while flushing record");
 316         }
 317
 318         --record->flush_group->refs;
 319         record->flush_group = NULL;
 320
 321         /*
 322          * Adjust the flush state and dependancy based on success or
 323          * failure.
 324          */
 325         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
 326                 if ((target_ip = record->target_ip) != NULL) {
 327                         TAILQ_REMOVE(&target_ip->target_list, record,
 328                                      target_entry);
 329                         record->target_ip = NULL;
 330                         hammer_test_inode(target_ip);
 331                 }
 332                 record->flush_state = HAMMER_FST_IDLE;
 333         } else {
 334                 if (record->target_ip) {
 335                         record->flush_state = HAMMER_FST_SETUP;
 336                         hammer_test_inode(record->ip);
 337                         hammer_test_inode(record->target_ip);
 338                 } else {
 339                         record->flush_state = HAMMER_FST_IDLE;
 340                 }
 341         }
 342         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 343
 344         /*
 345          * Cleanup
 346          */
 347         if (record->flags & HAMMER_RECF_WANTED) {
 348                 record->flags &= ~HAMMER_RECF_WANTED;
 349                 wakeup(record);
 350         }
 351         hammer_rel_mem_record(record);
 352 }
 353
 354 /*
 355  * Release a memory record.  Records marked for deletion are immediately
 356  * removed from the RB-Tree but otherwise left intact until the last ref
 357  * goes away.
 358  */
 359 void
 360 hammer_rel_mem_record(struct hammer_record *record)
 361 {
 362         hammer_mount_t hmp;
 363         hammer_reserve_t resv;
 364         hammer_inode_t ip;
 365         hammer_inode_t target_ip;
 366
 367         hammer_unref(&record->lock);
 368
 369         if (record->lock.refs == 0) {
 370                 /*
 371                  * Upon release of the last reference wakeup any waiters.
 372                  * The record structure may get destroyed so callers will
 373                  * loop up and do a relookup.
 374                  *
 375                  * WARNING!  Record must be removed from RB-TREE before we
 376                  * might possibly block.  hammer_test_inode() can block!
 377                  */
 378                 ip = record->ip;
 379                 hmp = ip->hmp;
 380
 381                 /*
 382                  * Upon release of the last reference a record marked deleted
 383                  * by the front or backend, or committed by the backend,
 384                  * is destroyed.
 385                  */
 386                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 387                                      HAMMER_RECF_DELETED_BE |
 388                                      HAMMER_RECF_COMMITTED)) {
 389                         KKASSERT(ip->lock.refs > 0);
 390                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 391
 392                         /*
 393                          * target_ip may have zero refs, we have to ref it
 394                          * to prevent it from being ripped out from under
 395                          * us.
 396                          */
 397                         if ((target_ip = record->target_ip) != NULL) {
 398                                 TAILQ_REMOVE(&target_ip->target_list,
 399                                              record, target_entry);
 400                                 record->target_ip = NULL;
 401                                 hammer_ref(&target_ip->lock);
 402                         }
 403
 404                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 405                                 RB_REMOVE(hammer_rec_rb_tree,
 406                                           &record->ip->rec_tree,
 407                                           record);
 408                                 KKASSERT(ip->rsv_recs > 0);
 409                                 --hmp->rsv_recs;
 410                                 --ip->rsv_recs;
 411                                 hmp->rsv_databytes -= record->leaf.data_len;
 412                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 413
 414                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 415                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 416                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 417                                         hammer_test_inode(record->ip);
 418                                 }
 419                         }
 420
 421                         /*
 422                          * We must wait for any direct-IO to complete before
 423                          * we can destroy the record because the bio may
 424                          * have a reference to it.
 425                          */
 426                         if (record->flags &
 427                            (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL)) {
 428                                 hammer_io_direct_wait(record);
 429                         }
 430
 431
 432                         /*
 433                          * Do this test after removing record from the B-Tree.
 434                          */
 435                         if (target_ip) {
 436                                 hammer_test_inode(target_ip);
 437                                 hammer_rel_inode(target_ip, 0);
 438                         }
 439
 440                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 441                                 --hammer_count_record_datas;
 442                                 kfree(record->data, hmp->m_misc);
 443                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 444                         }
 445
 446                         /*
 447                          * Release the reservation.
 448                          *
 449                          * If the record was not committed we can theoretically
 450                          * undo the reservation.  However, doing so might
 451                          * create weird edge cases with the ordering of
 452                          * direct writes because the related buffer cache
 453                          * elements are per-vnode.  So we don't try.
 454                          */
 455                         if ((resv = record->resv) != NULL) {
 456                                 /* XXX undo leaf.data_offset,leaf.data_len */
 457                                 hammer_blockmap_reserve_complete(hmp, resv);
 458                                 record->resv = NULL;
 459                         }
 460                         record->data = NULL;
 461                         --hammer_count_records;
 462                         kfree(record, hmp->m_misc);
 463                 }
 464         }
 465 }
 466
 467 /*
 468  * Record visibility depends on whether the record is being accessed by
 469  * the backend or the frontend.  Backend tests ignore the frontend delete
 470  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 471  * must also check for commit races.
 472  *
 473  * Return non-zero if the record is visible, zero if it isn't or if it is
 474  * deleted.  Returns 0 if the record has been comitted (unless the special
 475  * delete-visibility flag is set).  A committed record must be located
 476  * via the media B-Tree.  Returns non-zero if the record is good.
 477  *
 478  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 479  * records to be returned.  This is so pending deletions are detected
 480  * when using an iterator to locate an unused hash key, or when we need
 481  * to locate historical records on-disk to destroy.
 482  */
 483 static __inline
 484 int
 485 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 486 {
 487         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 488                 return(1);
 489         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 490                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 491                                      HAMMER_RECF_COMMITTED)) {
 492                         return(0);
 493                 }
 494         } else {
 495                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 496                                      HAMMER_RECF_DELETED_BE |
 497                                      HAMMER_RECF_COMMITTED)) {
 498                         return(0);
 499                 }
 500         }
 501         return(1);
 502 }
 503
 504 /*
 505  * This callback is used as part of the RB_SCAN function for in-memory
 506  * records.  We terminate it (return -1) as soon as we get a match.
 507  *
 508  * This routine is used by frontend code.
 509  *
 510  * The primary compare code does not account for ASOF lookups.  This
 511  * code handles that case as well as a few others.
 512  */
 513 static
 514 int
 515 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 516 {
 517         hammer_cursor_t cursor = data;
 518
 519         /*
 520          * We terminate on success, so this should be NULL on entry.
 521          */
 522         KKASSERT(cursor->iprec == NULL);
 523
 524         /*
 525          * Skip if the record was marked deleted or committed.
 526          */
 527         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 528                 return(0);
 529
 530         /*
 531          * Skip if not visible due to our as-of TID
 532          */
 533         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 534                 if (cursor->asof < rec->leaf.base.create_tid)
 535                         return(0);
 536                 if (rec->leaf.base.delete_tid &&
 537                     cursor->asof >= rec->leaf.base.delete_tid) {
 538                         return(0);
 539                 }
 540         }
 541
 542         /*
 543          * ref the record.  The record is protected from backend B-Tree
 544          * interactions by virtue of the cursor's IP lock.
 545          */
 546         hammer_ref(&rec->lock);
 547
 548         /*
 549          * The record may have been deleted or committed while we
 550          * were blocked.  XXX remove?
 551          */
 552         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 553                 hammer_rel_mem_record(rec);
 554                 return(0);
 555         }
 556
 557         /*
 558          * Set the matching record and stop the scan.
 559          */
 560         cursor->iprec = rec;
 561         return(-1);
 562 }
 563
 564
 565 /*
 566  * Lookup an in-memory record given the key specified in the cursor.  Works
 567  * just like hammer_btree_lookup() but operates on an inode's in-memory
 568  * record list.
 569  *
 570  * The lookup must fail if the record is marked for deferred deletion.
 571  *
 572  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 573  */
 574 static
 575 int
 576 hammer_mem_lookup(hammer_cursor_t cursor)
 577 {
 578         KKASSERT(cursor->ip);
 579         if (cursor->iprec) {
 580                 hammer_rel_mem_record(cursor->iprec);
 581                 cursor->iprec = NULL;
 582         }
 583         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 584                                    hammer_rec_scan_callback, cursor);
 585
 586         return (cursor->iprec ? 0 : ENOENT);
 587 }
 588
 589 /*
 590  * hammer_mem_first() - locate the first in-memory record matching the
 591  * cursor within the bounds of the key range.
 592  *
 593  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 594  * will set ATEMEM the same as MEMEOF, and does not return any error.
 595  */
 596 static
 597 void
 598 hammer_mem_first(hammer_cursor_t cursor)
 599 {
 600         hammer_inode_t ip;
 601
 602         ip = cursor->ip;
 603         KKASSERT(ip != NULL);
 604
 605         if (cursor->iprec) {
 606                 hammer_rel_mem_record(cursor->iprec);
 607                 cursor->iprec = NULL;
 608         }
 609         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 610                                    hammer_rec_scan_callback, cursor);
 611
 612         if (cursor->iprec)
 613                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
 614         else
 615                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
 616 }
 617
 618 /************************************************************************
 619  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 620  ************************************************************************
 621  *
 622  * These functions manipulate in-memory records.  Such records typically
 623  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 624  */
 625
 626 /*
 627  * Add a directory entry (dip,ncp) which references inode (ip).
 628  *
 629  * Note that the low 32 bits of the namekey are set temporarily to create
 630  * a unique in-memory record, and may be modified a second time when the
 631  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 632  * all 0's when synching to disk, which is not handled here.
 633  *
 634  * NOTE: bytes does not include any terminating \0 on name, and name might
 635  * not be terminated.
 636  */
 637 int
 638 hammer_ip_add_directory(struct hammer_transaction *trans,
 639                      struct hammer_inode *dip, const char *name, int bytes,
 640                      struct hammer_inode *ip)
 641 {
 642         struct hammer_cursor cursor;
 643         hammer_record_t record;
 644         int error;
 645         u_int32_t max_iterations;
 646
 647         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 648
 649         record->type = HAMMER_MEM_RECORD_ADD;
 650         record->leaf.base.localization = dip->obj_localization +
 651                                          hammer_dir_localization(dip);
 652         record->leaf.base.obj_id = dip->obj_id;
 653         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 654                                                          &max_iterations);
 655         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 656         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 657         record->data->entry.obj_id = ip->obj_id;
 658         record->data->entry.localization = ip->obj_localization;
 659         bcopy(name, record->data->entry.name, bytes);
 660
 661         ++ip->ino_data.nlinks;
 662         ip->ino_data.ctime = trans->time;
 663         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 664
 665         /*
 666          * Find an unused namekey.  Both the in-memory record tree and
 667          * the B-Tree are checked.  We do not want historically deleted
 668          * names to create a collision as our iteration space may be limited,
 669          * and since create_tid wouldn't match anyway an ASOF search
 670          * must be used to locate collisions.
 671          *
 672          * delete-visibility is set so pending deletions do not give us
 673          * a false-negative on our ability to use an iterator.
 674          *
 675          * The iterator must not rollover the key.  Directory keys only
 676          * use the positive key space.
 677          */
 678         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 679         cursor.key_beg = record->leaf.base;
 680         cursor.flags |= HAMMER_CURSOR_ASOF;
 681         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 682         cursor.asof = ip->obj_asof;
 683
 684         while (hammer_ip_lookup(&cursor) == 0) {
 685                 ++record->leaf.base.key;
 686                 KKASSERT(record->leaf.base.key > 0);
 687                 cursor.key_beg.key = record->leaf.base.key;
 688                 if (--max_iterations == 0) {
 689                         hammer_rel_mem_record(record);
 690                         error = ENOSPC;
 691                         goto failed;
 692                 }
 693         }
 694
 695         /*
 696          * The target inode and the directory entry are bound together.
 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 the idle
 704          * state.  An inode not in an idle state is given an extra reference.
 705          *
 706          * When transitioning to a SETUP state flag for an automatic reflush
 707          * when the dependancies are disposed of if someone is waiting on
 708          * the inode.
 709          */
 710         if (ip->flush_state == HAMMER_FST_IDLE) {
 711                 hammer_ref(&ip->lock);
 712                 ip->flush_state = HAMMER_FST_SETUP;
 713                 if (ip->flags & HAMMER_INODE_FLUSHW)
 714                         ip->flags |= HAMMER_INODE_REFLUSH;
 715         }
 716         error = hammer_mem_add(record);
 717         if (error == 0) {
 718                 dip->ino_data.mtime = trans->time;
 719                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 720         }
 721 failed:
 722         hammer_done_cursor(&cursor);
 723         return(error);
 724 }
 725
 726 /*
 727  * Delete the directory entry and update the inode link count.  The
 728  * cursor must be seeked to the directory entry record being deleted.
 729  *
 730  * The related inode should be share-locked by the caller.  The caller is
 731  * on the frontend.  It could also be NULL indicating that the directory
 732  * entry being removed has no related inode.
 733  *
 734  * This function can return EDEADLK requiring the caller to terminate
 735  * the cursor, any locks, wait on the returned record, and retry.
 736  */
 737 int
 738 hammer_ip_del_directory(struct hammer_transaction *trans,
 739                      hammer_cursor_t cursor, struct hammer_inode *dip,
 740                      struct hammer_inode *ip)
 741 {
 742         hammer_record_t record;
 743         int error;
 744
 745         if (hammer_cursor_inmem(cursor)) {
 746                 /*
 747                  * In-memory (unsynchronized) records can simply be freed.
 748                  *
 749                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 750                  * by the backend, we must still avoid races against the
 751                  * backend potentially syncing the record to the media.
 752                  *
 753                  * We cannot call hammer_ip_delete_record(), that routine may
 754                  * only be called from the backend.
 755                  */
 756                 record = cursor->iprec;
 757                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 758                                      HAMMER_RECF_DELETED_BE |
 759                                      HAMMER_RECF_COMMITTED)) {
 760                         KKASSERT(cursor->deadlk_rec == NULL);
 761                         hammer_ref(&record->lock);
 762                         cursor->deadlk_rec = record;
 763                         error = EDEADLK;
 764                 } else {
 765                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 766                         record->flags |= HAMMER_RECF_DELETED_FE;
 767                         error = 0;
 768                 }
 769         } else {
 770                 /*
 771                  * If the record is on-disk we have to queue the deletion by
 772                  * the record's key.  This also causes lookups to skip the
 773                  * record.
 774                  */
 775                 KKASSERT(dip->flags &
 776                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 777                 record = hammer_alloc_mem_record(dip, 0);
 778                 record->type = HAMMER_MEM_RECORD_DEL;
 779                 record->leaf.base = cursor->leaf->base;
 780
 781                 /*
 782                  * ip may be NULL, indicating the deletion of a directory
 783                  * entry which has no related inode.
 784                  */
 785                 record->target_ip = ip;
 786                 if (ip) {
 787                         record->flush_state = HAMMER_FST_SETUP;
 788                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 789                                           target_entry);
 790                 } else {
 791                         record->flush_state = HAMMER_FST_IDLE;
 792                 }
 793
 794                 /*
 795                  * The inode now has a dependancy and must be taken out of
 796                  * the idle state.  An inode not in an idle state is given
 797                  * an extra reference.
 798                  *
 799                  * When transitioning to a SETUP state flag for an automatic
 800                  * reflush when the dependancies are disposed of if someone
 801                  * is waiting on the inode.
 802                  */
 803                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 804                         hammer_ref(&ip->lock);
 805                         ip->flush_state = HAMMER_FST_SETUP;
 806                         if (ip->flags & HAMMER_INODE_FLUSHW)
 807                                 ip->flags |= HAMMER_INODE_REFLUSH;
 808                 }
 809
 810                 error = hammer_mem_add(record);
 811         }
 812
 813         /*
 814          * One less link.  The file may still be open in the OS even after
 815          * all links have gone away.
 816          *
 817          * We have to terminate the cursor before syncing the inode to
 818          * avoid deadlocking against ourselves.  XXX this may no longer
 819          * be true.
 820          *
 821          * If nlinks drops to zero and the vnode is inactive (or there is
 822          * no vnode), call hammer_inode_unloadable_check() to zonk the
 823          * inode.  If we don't do this here the inode will not be destroyed
 824          * on-media until we unmount.
 825          */
 826         if (error == 0) {
 827                 if (ip) {
 828                         --ip->ino_data.nlinks;  /* do before we might block */
 829                         ip->ino_data.ctime = trans->time;
 830                 }
 831                 dip->ino_data.mtime = trans->time;
 832                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 833                 if (ip) {
 834                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 835                         if (ip->ino_data.nlinks == 0 &&
 836                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 837                                 hammer_done_cursor(cursor);
 838                                 hammer_inode_unloadable_check(ip, 1);
 839                                 hammer_flush_inode(ip, 0);
 840                         }
 841                 }
 842
 843         }
 844         return(error);
 845 }
 846
 847 /*
 848  * Add a record to an inode.
 849  *
 850  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 851  * initialize the following additional fields:
 852  *
 853  * The related inode should be share-locked by the caller.  The caller is
 854  * on the frontend.
 855  *
 856  * record->rec.entry.base.base.key
 857  * record->rec.entry.base.base.rec_type
 858  * record->rec.entry.base.base.data_len
 859  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 860  */
 861 int
 862 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 863 {
 864         hammer_inode_t ip = record->ip;
 865         int error;
 866
 867         KKASSERT(record->leaf.base.localization != 0);
 868         record->leaf.base.obj_id = ip->obj_id;
 869         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 870         error = hammer_mem_add(record);
 871         return(error);
 872 }
 873
 874 /*
 875  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 876  * reserved so the front-end can flush large data writes without having
 877  * to queue the BIO to the flusher.  Only the related record gets queued
 878  * to the flusher.
 879  */
 880
 881 static hammer_record_t
 882 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 883 {
 884         struct hammer_bulk_info info;
 885
 886         bzero(&info, sizeof(info));
 887         info.leaf.base.obj_id = ip->obj_id;
 888         info.leaf.base.key = file_offset + bytes;
 889         info.leaf.base.create_tid = 0;
 890         info.leaf.base.delete_tid = 0;
 891         info.leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 892         info.leaf.base.obj_type = 0;                            /* unused */
 893         info.leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;        /* unused */
 894         info.leaf.base.localization = ip->obj_localization +    /* unused */
 895                                       HAMMER_LOCALIZE_MISC;
 896         info.leaf.data_len = bytes;
 897
 898         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 899                                    hammer_bulk_scan_callback, &info);
 900
 901         return(info.record);    /* may be NULL */
 902 }
 903
 904 /*
 905  * Take records vetted by overlap_cmp.  The first non-deleted record
 906  * (if any) stops the scan.
 907  */
 908 static int
 909 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 910 {
 911         struct hammer_bulk_info *info = data;
 912
 913         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 914                              HAMMER_RECF_COMMITTED)) {
 915                 return(0);
 916         }
 917         hammer_ref(&record->lock);
 918         info->record = record;
 919         return(-1);                     /* stop scan */
 920 }
 921
 922 /*
 923  * Reserve blockmap space placemarked with an in-memory record.
 924  *
 925  * This routine is called by the frontend in order to be able to directly
 926  * flush a buffer cache buffer.  The frontend has locked the related buffer
 927  * cache buffers and we should be able to manipulate any overlapping
 928  * in-memory records.
 929  *
 930  * The caller is responsible for adding the returned record.
 931  */
 932 hammer_record_t
 933 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 934                    int *errorp)
 935 {
 936         hammer_record_t record;
 937         hammer_record_t conflict;
 938         int zone;
 939
 940         /*
 941          * Deal with conflicting in-memory records.  We cannot have multiple
 942          * in-memory records for the same base offset without seriously
 943          * confusing the backend, including but not limited to the backend
 944          * issuing delete-create-delete or create-delete-create sequences
 945          * and asserting on the delete_tid being the same as the create_tid.
 946          *
 947          * If we encounter a record with the backend interlock set we cannot
 948          * immediately delete it without confusing the backend.
 949          */
 950         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 951                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 952                         conflict->flags |= HAMMER_RECF_WANTED;
 953                         tsleep(conflict, 0, "hmrrc3", 0);
 954                 } else {
 955                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 956                 }
 957                 hammer_rel_mem_record(conflict);
 958         }
 959
 960         /*
 961          * Create a record to cover the direct write.  This is called with
 962          * the related BIO locked so there should be no possible conflict.
 963          *
 964          * The backend is responsible for finalizing the space reserved in
 965          * this record.
 966          *
 967          * XXX bytes not aligned, depend on the reservation code to
 968          * align the reservation.
 969          */
 970         record = hammer_alloc_mem_record(ip, 0);
 971         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 972                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 973         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 974                                                &record->leaf.data_offset,
 975                                                errorp);
 976         if (record->resv == NULL) {
 977                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 978                 hammer_rel_mem_record(record);
 979                 return(NULL);
 980         }
 981         record->type = HAMMER_MEM_RECORD_DATA;
 982         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 983         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 984         record->leaf.base.obj_id = ip->obj_id;
 985         record->leaf.base.key = file_offset + bytes;
 986         record->leaf.base.localization = ip->obj_localization +
 987                                          HAMMER_LOCALIZE_MISC;
 988         record->leaf.data_len = bytes;
 989         hammer_crc_set_leaf(data, &record->leaf);
 990         KKASSERT(*errorp == 0);
 991         return(record);
 992 }
 993
 994 /*
 995  * Frontend truncation code.  Scan in-memory records only.  On-disk records
 996  * and records in a flushing state are handled by the backend.  The vnops
 997  * setattr code will handle the block containing the truncation point.
 998  *
 999  * Partial blocks are not deleted.
1000  */
1001 int
1002 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1003 {
1004         struct rec_trunc_info info;
1005
1006         switch(ip->ino_data.obj_type) {
1007         case HAMMER_OBJTYPE_REGFILE:
1008                 info.rec_type = HAMMER_RECTYPE_DATA;
1009                 break;
1010         case HAMMER_OBJTYPE_DBFILE:
1011                 info.rec_type = HAMMER_RECTYPE_DB;
1012                 break;
1013         default:
1014                 return(EINVAL);
1015         }
1016         info.trunc_off = file_size;
1017         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1018                                    hammer_frontend_trunc_callback, &info);
1019         return(0);
1020 }
1021
1022 static int
1023 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1024 {
1025         if (record->flags & HAMMER_RECF_DELETED_FE)
1026                 return(0);
1027         if (record->flush_state == HAMMER_FST_FLUSH)
1028                 return(0);
1029         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1030         hammer_ref(&record->lock);
1031         record->flags |= HAMMER_RECF_DELETED_FE;
1032         hammer_rel_mem_record(record);
1033         return(0);
1034 }
1035
1036 /*
1037  * Return 1 if the caller must check for and delete existing records
1038  * before writing out a new data record.
1039  *
1040  * Return 0 if the caller can just insert the record into the B-Tree without
1041  * checking.
1042  */
1043 static int
1044 hammer_record_needs_overwrite_delete(hammer_record_t record)
1045 {
1046         hammer_inode_t ip = record->ip;
1047         int64_t file_offset;
1048         int r;
1049
1050         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1051                 file_offset = record->leaf.base.key;
1052         else
1053                 file_offset = record->leaf.base.key - record->leaf.data_len;
1054         r = (file_offset < ip->save_trunc_off);
1055         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1056                 if (ip->save_trunc_off <= record->leaf.base.key)
1057                         ip->save_trunc_off = record->leaf.base.key + 1;
1058         } else {
1059                 if (ip->save_trunc_off < record->leaf.base.key)
1060                         ip->save_trunc_off = record->leaf.base.key;
1061         }
1062         return(r);
1063 }
1064
1065 /*
1066  * Backend code.  Sync a record to the media.
1067  */
1068 int
1069 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1070 {
1071         hammer_transaction_t trans = cursor->trans;
1072         int64_t file_offset;
1073         int bytes;
1074         void *bdata;
1075         int error;
1076         int doprop;
1077
1078         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1079         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1080         KKASSERT(record->leaf.base.localization != 0);
1081
1082         /*
1083          * Any direct-write related to the record must complete before we
1084          * can sync the record to the on-disk media.
1085          */
1086         if (record->flags & (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL))
1087                 hammer_io_direct_wait(record);
1088
1089         /*
1090          * If this is a bulk-data record placemarker there may be an existing
1091          * record on-disk, indicating a data overwrite.  If there is the
1092          * on-disk record must be deleted before we can insert our new record.
1093          *
1094          * We've synthesized this record and do not know what the create_tid
1095          * on-disk is, nor how much data it represents.
1096          *
1097          * Keep in mind that (key) for data records is (base_offset + len),
1098          * not (base_offset).  Also, we only want to get rid of on-disk
1099          * records since we are trying to sync our in-memory record, call
1100          * hammer_ip_delete_range() with truncating set to 1 to make sure
1101          * it skips in-memory records.
1102          *
1103          * It is ok for the lookup to return ENOENT.
1104          *
1105          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1106          * to call hammer_ip_delete_range() or not.  This also means we must
1107          * update sync_trunc_off() as we write.
1108          */
1109         if (record->type == HAMMER_MEM_RECORD_DATA &&
1110             hammer_record_needs_overwrite_delete(record)) {
1111                 file_offset = record->leaf.base.key - record->leaf.data_len;
1112                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1113                         ~HAMMER_BUFMASK;
1114                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1115                 error = hammer_ip_delete_range(
1116                                 cursor, record->ip,
1117                                 file_offset, file_offset + bytes - 1,
1118                                 1);
1119                 if (error && error != ENOENT)
1120                         goto done;
1121         }
1122
1123         /*
1124          * If this is a general record there may be an on-disk version
1125          * that must be deleted before we can insert the new record.
1126          */
1127         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1128                 error = hammer_delete_general(cursor, record->ip,
1129                                               &record->leaf);
1130                 if (error && error != ENOENT)
1131                         goto done;
1132         }
1133
1134         /*
1135          * Setup the cursor.
1136          */
1137         hammer_normalize_cursor(cursor);
1138         cursor->key_beg = record->leaf.base;
1139         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1140         cursor->flags |= HAMMER_CURSOR_BACKEND;
1141         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1142
1143         /*
1144          * Records can wind up on-media before the inode itself is on-media.
1145          * Flag the case.
1146          */
1147         record->ip->flags |= HAMMER_INODE_DONDISK;
1148
1149         /*
1150          * If we are deleting a directory entry an exact match must be
1151          * found on-disk.
1152          */
1153         if (record->type == HAMMER_MEM_RECORD_DEL) {
1154                 error = hammer_btree_lookup(cursor);
1155                 if (error == 0) {
1156                         KKASSERT(cursor->iprec == NULL);
1157                         error = hammer_ip_delete_record(cursor, record->ip,
1158                                                         trans->tid);
1159                         if (error == 0) {
1160                                 record->flags |= HAMMER_RECF_DELETED_BE |
1161                                                  HAMMER_RECF_COMMITTED;
1162                                 ++record->ip->rec_generation;
1163                         }
1164                 }
1165                 goto done;
1166         }
1167
1168         /*
1169          * We are inserting.
1170          *
1171          * Issue a lookup to position the cursor and locate the insertion
1172          * point.  The target key should not exist.  If we are creating a
1173          * directory entry we may have to iterate the low 32 bits of the
1174          * key to find an unused key.
1175          */
1176         hammer_sync_lock_sh(trans);
1177         cursor->flags |= HAMMER_CURSOR_INSERT;
1178         error = hammer_btree_lookup(cursor);
1179         if (hammer_debug_inode)
1180                 kprintf("DOINSERT LOOKUP %d\n", error);
1181         if (error == 0) {
1182                 kprintf("hammer_ip_sync_record: duplicate rec "
1183                         "at (%016llx)\n", (long long)record->leaf.base.key);
1184                 Debugger("duplicate record1");
1185                 error = EIO;
1186         }
1187 #if 0
1188         if (record->type == HAMMER_MEM_RECORD_DATA)
1189                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1190                         record->leaf.base.key - record->leaf.data_len,
1191                         record->leaf.data_offset, error);
1192 #endif
1193
1194         if (error != ENOENT)
1195                 goto done_unlock;
1196
1197         /*
1198          * Allocate the record and data.  The result buffers will be
1199          * marked as being modified and further calls to
1200          * hammer_modify_buffer() will result in unneeded UNDO records.
1201          *
1202          * Support zero-fill records (data == NULL and data_len != 0)
1203          */
1204         if (record->type == HAMMER_MEM_RECORD_DATA) {
1205                 /*
1206                  * The data portion of a bulk-data record has already been
1207                  * committed to disk, we need only adjust the layer2
1208                  * statistics in the same transaction as our B-Tree insert.
1209                  */
1210                 KKASSERT(record->leaf.data_offset != 0);
1211                 error = hammer_blockmap_finalize(trans,
1212                                                  record->resv,
1213                                                  record->leaf.data_offset,
1214                                                  record->leaf.data_len);
1215         } else if (record->data && record->leaf.data_len) {
1216                 /*
1217                  * Wholely cached record, with data.  Allocate the data.
1218                  */
1219                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1220                                           record->leaf.base.rec_type,
1221                                           &record->leaf.data_offset,
1222                                           &cursor->data_buffer,
1223                                           0, &error);
1224                 if (bdata == NULL)
1225                         goto done_unlock;
1226                 hammer_crc_set_leaf(record->data, &record->leaf);
1227                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1228                 bcopy(record->data, bdata, record->leaf.data_len);
1229                 hammer_modify_buffer_done(cursor->data_buffer);
1230         } else {
1231                 /*
1232                  * Wholely cached record, without data.
1233                  */
1234                 record->leaf.data_offset = 0;
1235                 record->leaf.data_crc = 0;
1236         }
1237
1238         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1239         if (hammer_debug_inode && error) {
1240                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1241                         error,
1242                         (long long)cursor->node->node_offset,
1243                         cursor->index,
1244                         (long long)record->leaf.base.key);
1245         }
1246
1247         /*
1248          * Our record is on-disk and we normally mark the in-memory version
1249          * as having been committed (and not BE-deleted).
1250          *
1251          * If the record represented a directory deletion but we had to
1252          * sync a valid directory entry to disk due to dependancies,
1253          * we must convert the record to a covering delete so the
1254          * frontend does not have visibility on the synced entry.
1255          */
1256         if (error == 0) {
1257                 if (doprop) {
1258                         hammer_btree_do_propagation(cursor,
1259                                                     record->ip->pfsm,
1260                                                     &record->leaf);
1261                 }
1262                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1263                         /*
1264                          * Must convert deleted directory entry add
1265                          * to a directory entry delete.
1266                          */
1267                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1268                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1269                         record->type = HAMMER_MEM_RECORD_DEL;
1270                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1271                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1272                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1273                                                  HAMMER_RECF_DELETED_BE)) == 0);
1274                         /* converted record is not yet committed */
1275                         /* hammer_flush_record_done takes care of the rest */
1276                 } else {
1277                         /*
1278                          * Everything went fine and we are now done with
1279                          * this record.
1280                          */
1281                         record->flags |= HAMMER_RECF_COMMITTED;
1282                         ++record->ip->rec_generation;
1283                 }
1284         } else {
1285                 if (record->leaf.data_offset) {
1286                         hammer_blockmap_free(trans, record->leaf.data_offset,
1287                                              record->leaf.data_len);
1288                 }
1289         }
1290 done_unlock:
1291         hammer_sync_unlock(trans);
1292 done:
1293         return(error);
1294 }
1295
1296 /*
1297  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1298  * entry's key is used to deal with hash collisions in the upper 32 bits.
1299  * A unique 64 bit key is generated in-memory and may be regenerated a
1300  * second time when the directory record is flushed to the on-disk B-Tree.
1301  *
1302  * A referenced record is passed to this function.  This function
1303  * eats the reference.  If an error occurs the record will be deleted.
1304  *
1305  * A copy of the temporary record->data pointer provided by the caller
1306  * will be made.
1307  */
1308 int
1309 hammer_mem_add(hammer_record_t record)
1310 {
1311         hammer_mount_t hmp = record->ip->hmp;
1312
1313         /*
1314          * Make a private copy of record->data
1315          */
1316         if (record->data)
1317                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1318
1319         /*
1320          * Insert into the RB tree.  A unique key should have already
1321          * been selected if this is a directory entry.
1322          */
1323         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1324                 record->flags |= HAMMER_RECF_DELETED_FE;
1325                 hammer_rel_mem_record(record);
1326                 return (EEXIST);
1327         }
1328         ++hmp->count_newrecords;
1329         ++hmp->rsv_recs;
1330         ++record->ip->rsv_recs;
1331         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1332         record->flags |= HAMMER_RECF_ONRBTREE;
1333         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1334         hammer_rel_mem_record(record);
1335         return(0);
1336 }
1337
1338 /************************************************************************
1339  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1340  ************************************************************************
1341  *
1342  * These functions augment the B-Tree scanning functions in hammer_btree.c
1343  * by merging in-memory records with on-disk records.
1344  */
1345
1346 /*
1347  * Locate a particular record either in-memory or on-disk.
1348  *
1349  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1350  * NOT be called to iterate results.
1351  */
1352 int
1353 hammer_ip_lookup(hammer_cursor_t cursor)
1354 {
1355         int error;
1356
1357         /*
1358          * If the element is in-memory return it without searching the
1359          * on-disk B-Tree
1360          */
1361         KKASSERT(cursor->ip);
1362         error = hammer_mem_lookup(cursor);
1363         if (error == 0) {
1364                 cursor->leaf = &cursor->iprec->leaf;
1365                 return(error);
1366         }
1367         if (error != ENOENT)
1368                 return(error);
1369
1370         /*
1371          * If the inode has on-disk components search the on-disk B-Tree.
1372          */
1373         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1374                 return(error);
1375         error = hammer_btree_lookup(cursor);
1376         if (error == 0)
1377                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1378         return(error);
1379 }
1380
1381 /*
1382  * Helper for hammer_ip_first()/hammer_ip_next()
1383  *
1384  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1385  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1386  * state if hammer_ip_next() needs to re-seek.
1387  */
1388 static __inline
1389 int
1390 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1391 {
1392         hammer_inode_t ip = cursor->ip;
1393         int error;
1394
1395         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1396                 error = hammer_btree_lookup(cursor);
1397                 if (error == ENOENT || error == EDEADLK) {
1398                         if (hammer_debug_general & 0x2000) {
1399                                 kprintf("error %d node %p %016llx index %d\n",
1400                                         error, cursor->node,
1401                                         (long long)cursor->node->node_offset,
1402                                         cursor->index);
1403                         }
1404                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1405                         error = hammer_btree_iterate(cursor);
1406                 }
1407                 if (error == 0) {
1408                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1409                                            HAMMER_CURSOR_ATEDISK);
1410                 } else {
1411                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1412                                          HAMMER_CURSOR_ATEDISK;
1413                         if (error == ENOENT)
1414                                 error = 0;
1415                 }
1416         } else {
1417                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1418                 error = 0;
1419         }
1420         return(error);
1421 }
1422
1423 /*
1424  * Helper for hammer_ip_next()
1425  *
1426  * The caller has determined that the media cursor is further along than the
1427  * memory cursor and must be reseeked after a generation number change.
1428  */
1429 static
1430 int
1431 _hammer_ip_reseek(hammer_cursor_t cursor)
1432 {
1433         struct hammer_base_elm save;
1434         hammer_btree_elm_t elm;
1435         int error;
1436         int r;
1437         int again = 0;
1438
1439         /*
1440          * Do the re-seek.
1441          */
1442         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1443                 (long long)cursor->ip->obj_id);
1444         save = cursor->key_beg;
1445         cursor->key_beg = cursor->iprec->leaf.base;
1446         error = _hammer_ip_seek_btree(cursor);
1447         KKASSERT(error == 0);
1448         cursor->key_beg = save;
1449
1450         /*
1451          * If the memory record was previous returned to
1452          * the caller and the media record matches
1453          * (-1/+1: only create_tid differs), then iterate
1454          * the media record to avoid a double result.
1455          */
1456         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1457             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1458                 elm = &cursor->node->ondisk->elms[cursor->index];
1459                 r = hammer_btree_cmp(&elm->base,
1460                                      &cursor->iprec->leaf.base);
1461                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1462                         if (r >= -1 && r <= 1) {
1463                                 kprintf("HAMMER: Debug: iterated after "
1464                                         "re-seek (asof r=%d)\n", r);
1465                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1466                                 again = 1;
1467                         }
1468                 } else {
1469                         if (r == 0) {
1470                                 kprintf("HAMMER: Debug: iterated after "
1471                                         "re-seek\n");
1472                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1473                                 again = 1;
1474                         }
1475                 }
1476         }
1477         return(again);
1478 }
1479
1480 /*
1481  * Locate the first record within the cursor's key_beg/key_end range,
1482  * restricted to a particular inode.  0 is returned on success, ENOENT
1483  * if no records matched the requested range, or some other error.
1484  *
1485  * When 0 is returned hammer_ip_next() may be used to iterate additional
1486  * records within the requested range.
1487  *
1488  * This function can return EDEADLK, requiring the caller to terminate
1489  * the cursor and try again.
1490  */
1491
1492 int
1493 hammer_ip_first(hammer_cursor_t cursor)
1494 {
1495         hammer_inode_t ip = cursor->ip;
1496         int error;
1497
1498         KKASSERT(ip != NULL);
1499
1500         /*
1501          * Clean up fields and setup for merged scan
1502          */
1503         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1504
1505         /*
1506          * Search the in-memory record list (Red-Black tree).  Unlike the
1507          * B-Tree search, mem_first checks for records in the range.
1508          *
1509          * This function will setup both ATEMEM and MEMEOF properly for
1510          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1511          */
1512         hammer_mem_first(cursor);
1513
1514         /*
1515          * Detect generation changes during blockages, including
1516          * blockages which occur on the initial btree search.
1517          */
1518         cursor->rec_generation = cursor->ip->rec_generation;
1519
1520         /*
1521          * Initial search and result
1522          */
1523         error = _hammer_ip_seek_btree(cursor);
1524         if (error == 0)
1525                 error = hammer_ip_next(cursor);
1526
1527         return (error);
1528 }
1529
1530 /*
1531  * Retrieve the next record in a merged iteration within the bounds of the
1532  * cursor.  This call may be made multiple times after the cursor has been
1533  * initially searched with hammer_ip_first().
1534  *
1535  * There are numerous special cases in this code to deal with races between
1536  * in-memory records and on-media records.
1537  *
1538  * 0 is returned on success, ENOENT if no further records match the
1539  * requested range, or some other error code is returned.
1540  */
1541 int
1542 hammer_ip_next(hammer_cursor_t cursor)
1543 {
1544         hammer_btree_elm_t elm;
1545         hammer_record_t rec;
1546         hammer_record_t tmprec;
1547         int error;
1548         int r;
1549
1550 again:
1551         /*
1552          * Get the next on-disk record
1553          *
1554          * NOTE: If we deleted the last on-disk record we had scanned
1555          *       ATEDISK will be clear and RETEST will be set, forcing
1556          *       a call to iterate.  The fact that ATEDISK is clear causes
1557          *       iterate to re-test the 'current' element.  If ATEDISK is
1558          *       set, iterate will skip the 'current' element.
1559          */
1560         error = 0;
1561         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1562                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1563                                      HAMMER_CURSOR_RETEST)) {
1564                         error = hammer_btree_iterate(cursor);
1565                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1566                         if (error == 0) {
1567                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1568                                 hammer_cache_node(&cursor->ip->cache[1],
1569                                                   cursor->node);
1570                         } else if (error == ENOENT) {
1571                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1572                                                  HAMMER_CURSOR_ATEDISK;
1573                                 error = 0;
1574                         }
1575                 }
1576         }
1577
1578         /*
1579          * If the generation changed the backend has deleted or committed
1580          * one or more memory records since our last check.
1581          *
1582          * When this case occurs if the disk cursor is > current memory record
1583          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1584          * Since the cursor is ahead it must have not yet been eaten (if
1585          * not at eof anyway). (XXX data offset case?)
1586          *
1587          * NOTE: we are not doing a full check here.  That will be handled
1588          * later on.
1589          *
1590          * If we have exhausted all memory records we do not have to do any
1591          * further seeks.
1592          */
1593         while (cursor->rec_generation != cursor->ip->rec_generation &&
1594                error == 0
1595         ) {
1596                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1597                 cursor->rec_generation = cursor->ip->rec_generation;
1598                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1599                         break;
1600                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1601                         r = 1;
1602                 } else {
1603                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1604                         elm = &cursor->node->ondisk->elms[cursor->index];
1605                         r = hammer_btree_cmp(&elm->base,
1606                                              &cursor->iprec->leaf.base);
1607                 }
1608
1609                 /*
1610                  * Do we re-seek the media cursor?
1611                  */
1612                 if (r > 0) {
1613                         if (_hammer_ip_reseek(cursor))
1614                                 goto again;
1615                 }
1616         }
1617
1618         /*
1619          * We can now safely get the next in-memory record.  We cannot
1620          * block here.
1621          *
1622          * hammer_rec_scan_cmp:  Is the record still in our general range,
1623          *                       (non-inclusive of snapshot exclusions)?
1624          * hammer_rec_scan_callback: Is the record in our snapshot?
1625          */
1626         tmprec = NULL;
1627         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1628                 /*
1629                  * If the current memory record was eaten then get the next
1630                  * one.  Stale records are skipped.
1631                  */
1632                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1633                         tmprec = cursor->iprec;
1634                         cursor->iprec = NULL;
1635                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1636                         while (rec) {
1637                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1638                                         break;
1639                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1640                                         break;
1641                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1642                         }
1643                         if (cursor->iprec) {
1644                                 KKASSERT(cursor->iprec == rec);
1645                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1646                         } else {
1647                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1648                         }
1649                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1650                 }
1651         }
1652
1653         /*
1654          * MEMORY RECORD VALIDITY TEST
1655          *
1656          * (We still can't block, which is why tmprec is being held so
1657          * long).
1658          *
1659          * If the memory record is no longer valid we skip it.  It may
1660          * have been deleted by the frontend.  If it was deleted or
1661          * committed by the backend the generation change re-seeked the
1662          * disk cursor and the record will be present there.
1663          */
1664         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1665                 KKASSERT(cursor->iprec);
1666                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1667                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1668                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1669                         if (tmprec)
1670                                 hammer_rel_mem_record(tmprec);
1671                         goto again;
1672                 }
1673         }
1674         if (tmprec)
1675                 hammer_rel_mem_record(tmprec);
1676
1677         /*
1678          * Extract either the disk or memory record depending on their
1679          * relative position.
1680          */
1681         error = 0;
1682         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1683         case 0:
1684                 /*
1685                  * Both entries valid.   Compare the entries and nominally
1686                  * return the first one in the sort order.  Numerous cases
1687                  * require special attention, however.
1688                  */
1689                 elm = &cursor->node->ondisk->elms[cursor->index];
1690                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1691
1692                 /*
1693                  * If the two entries differ only by their key (-2/2) or
1694                  * create_tid (-1/1), and are DATA records, we may have a
1695                  * nominal match.  We have to calculate the base file
1696                  * offset of the data.
1697                  */
1698                 if (r <= 2 && r >= -2 && r != 0 &&
1699                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1700                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1701                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1702                         int64_t base2 = cursor->iprec->leaf.base.key -
1703                                         cursor->iprec->leaf.data_len;
1704                         if (base1 == base2)
1705                                 r = 0;
1706                 }
1707
1708                 if (r < 0) {
1709                         error = hammer_btree_extract(cursor,
1710                                                      HAMMER_CURSOR_GET_LEAF);
1711                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1712                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1713                         break;
1714                 }
1715
1716                 /*
1717                  * If the entries match exactly the memory entry is either
1718                  * an on-disk directory entry deletion or a bulk data
1719                  * overwrite.  If it is a directory entry deletion we eat
1720                  * both entries.
1721                  *
1722                  * For the bulk-data overwrite case it is possible to have
1723                  * visibility into both, which simply means the syncer
1724                  * hasn't gotten around to doing the delete+insert sequence
1725                  * on the B-Tree.  Use the memory entry and throw away the
1726                  * on-disk entry.
1727                  *
1728                  * If the in-memory record is not either of these we
1729                  * probably caught the syncer while it was syncing it to
1730                  * the media.  Since we hold a shared lock on the cursor,
1731                  * the in-memory record had better be marked deleted at
1732                  * this point.
1733                  */
1734                 if (r == 0) {
1735                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1736                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1737                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1738                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1739                                         goto again;
1740                                 }
1741                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1742                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1743                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1744                                 }
1745                                 /* fall through to memory entry */
1746                         } else {
1747                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1748                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1749                                 goto again;
1750                         }
1751                 }
1752                 /* fall through to the memory entry */
1753         case HAMMER_CURSOR_ATEDISK:
1754                 /*
1755                  * Only the memory entry is valid.
1756                  */
1757                 cursor->leaf = &cursor->iprec->leaf;
1758                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1759                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1760
1761                 /*
1762                  * If the memory entry is an on-disk deletion we should have
1763                  * also had found a B-Tree record.  If the backend beat us
1764                  * to it it would have interlocked the cursor and we should
1765                  * have seen the in-memory record marked DELETED_FE.
1766                  */
1767                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1768                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1769                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1770                 }
1771                 break;
1772         case HAMMER_CURSOR_ATEMEM:
1773                 /*
1774                  * Only the disk entry is valid
1775                  */
1776                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1777                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1778                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1779                 break;
1780         default:
1781                 /*
1782                  * Neither entry is valid
1783                  *
1784                  * XXX error not set properly
1785                  */
1786                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1787                 cursor->leaf = NULL;
1788                 error = ENOENT;
1789                 break;
1790         }
1791         return(error);
1792 }
1793
1794 /*
1795  * Resolve the cursor->data pointer for the current cursor position in
1796  * a merged iteration.
1797  */
1798 int
1799 hammer_ip_resolve_data(hammer_cursor_t cursor)
1800 {
1801         hammer_record_t record;
1802         int error;
1803
1804         if (hammer_cursor_inmem(cursor)) {
1805                 /*
1806                  * The data associated with an in-memory record is usually
1807                  * kmalloced, but reserve-ahead data records will have an
1808                  * on-disk reference.
1809                  *
1810                  * NOTE: Reserve-ahead data records must be handled in the
1811                  * context of the related high level buffer cache buffer
1812                  * to interlock against async writes.
1813                  */
1814                 record = cursor->iprec;
1815                 cursor->data = record->data;
1816                 error = 0;
1817                 if (cursor->data == NULL) {
1818                         KKASSERT(record->leaf.base.rec_type ==
1819                                  HAMMER_RECTYPE_DATA);
1820                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1821                                                     record->leaf.data_offset,
1822                                                     record->leaf.data_len,
1823                                                     &error,
1824                                                     &cursor->data_buffer);
1825                 }
1826         } else {
1827                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1828                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1829         }
1830         return(error);
1831 }
1832
1833 /*
1834  * Backend truncation / record replacement - delete records in range.
1835  *
1836  * Delete all records within the specified range for inode ip.  In-memory
1837  * records still associated with the frontend are ignored.
1838  *
1839  * If truncating is non-zero in-memory records associated with the back-end
1840  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1841  *
1842  * NOTES:
1843  *
1844  *      * An unaligned range will cause new records to be added to cover
1845  *        the edge cases. (XXX not implemented yet).
1846  *
1847  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1848  *        also do not deal with unaligned ranges.
1849  *
1850  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1851  *
1852  *      * Record keys for regular file data have to be special-cased since
1853  *        they indicate the end of the range (key = base + bytes).
1854  *
1855  *      * This function may be asked to delete ridiculously huge ranges, for
1856  *        example if someone truncates or removes a 1TB regular file.  We
1857  *        must be very careful on restarts and we may have to stop w/
1858  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1859  */
1860 int
1861 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1862                        int64_t ran_beg, int64_t ran_end, int truncating)
1863 {
1864         hammer_transaction_t trans = cursor->trans;
1865         hammer_btree_leaf_elm_t leaf;
1866         int error;
1867         int64_t off;
1868         int64_t tmp64;
1869
1870 #if 0
1871         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1872 #endif
1873
1874         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1875 retry:
1876         hammer_normalize_cursor(cursor);
1877         cursor->key_beg.localization = ip->obj_localization +
1878                                        HAMMER_LOCALIZE_MISC;
1879         cursor->key_beg.obj_id = ip->obj_id;
1880         cursor->key_beg.create_tid = 0;
1881         cursor->key_beg.delete_tid = 0;
1882         cursor->key_beg.obj_type = 0;
1883
1884         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1885                 cursor->key_beg.key = ran_beg;
1886                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1887         } else {
1888                 /*
1889                  * The key in the B-Tree is (base+bytes), so the first possible
1890                  * matching key is ran_beg + 1.
1891                  */
1892                 cursor->key_beg.key = ran_beg + 1;
1893                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1894         }
1895
1896         cursor->key_end = cursor->key_beg;
1897         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1898                 cursor->key_end.key = ran_end;
1899         } else {
1900                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1901                 if (tmp64 < ran_end)
1902                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1903                 else
1904                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1905         }
1906
1907         cursor->asof = ip->obj_asof;
1908         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1909         cursor->flags |= HAMMER_CURSOR_ASOF;
1910         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1911         cursor->flags |= HAMMER_CURSOR_BACKEND;
1912         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1913
1914         error = hammer_ip_first(cursor);
1915
1916         /*
1917          * Iterate through matching records and mark them as deleted.
1918          */
1919         while (error == 0) {
1920                 leaf = cursor->leaf;
1921
1922                 KKASSERT(leaf->base.delete_tid == 0);
1923                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1924
1925                 /*
1926                  * There may be overlap cases for regular file data.  Also
1927                  * remember the key for a regular file record is (base + len),
1928                  * NOT (base).
1929                  *
1930                  * Note that do to duplicates (mem & media) allowed by
1931                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1932                  */
1933                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1934                         off = leaf->base.key - leaf->data_len;
1935                         /*
1936                          * Check the left edge case.  We currently do not
1937                          * split existing records.
1938                          */
1939                         if (off < ran_beg && leaf->base.key > ran_beg) {
1940                                 panic("hammer left edge case %016llx %d\n",
1941                                         (long long)leaf->base.key,
1942                                         leaf->data_len);
1943                         }
1944
1945                         /*
1946                          * Check the right edge case.  Note that the
1947                          * record can be completely out of bounds, which
1948                          * terminates the search.
1949                          *
1950                          * base->key is exclusive of the right edge while
1951                          * ran_end is inclusive of the right edge.  The
1952                          * (key - data_len) left boundary is inclusive.
1953                          *
1954                          * XXX theory-check this test at some point, are
1955                          * we missing a + 1 somewhere?  Note that ran_end
1956                          * could overflow.
1957                          */
1958                         if (leaf->base.key - 1 > ran_end) {
1959                                 if (leaf->base.key - leaf->data_len > ran_end)
1960                                         break;
1961                                 panic("hammer right edge case\n");
1962                         }
1963                 } else {
1964                         off = leaf->base.key;
1965                 }
1966
1967                 /*
1968                  * Delete the record.  When truncating we do not delete
1969                  * in-memory (data) records because they represent data
1970                  * written after the truncation.
1971                  *
1972                  * This will also physically destroy the B-Tree entry and
1973                  * data if the retention policy dictates.  The function
1974                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
1975                  * to retest the new 'current' element.
1976                  */
1977                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1978                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1979                         /*
1980                          * If we have built up too many meta-buffers we risk
1981                          * deadlocking the kernel and must stop.  This can
1982                          * occur when deleting ridiculously huge files.
1983                          * sync_trunc_off is updated so the next cycle does
1984                          * not re-iterate records we have already deleted.
1985                          *
1986                          * This is only done with formal truncations.
1987                          */
1988                         if (truncating > 1 && error == 0 &&
1989                             hammer_flusher_meta_limit(ip->hmp)) {
1990                                 ip->sync_trunc_off = off;
1991                                 error = EWOULDBLOCK;
1992                         }
1993                 }
1994                 if (error)
1995                         break;
1996                 ran_beg = off;  /* for restart */
1997                 error = hammer_ip_next(cursor);
1998         }
1999         if (cursor->node)
2000                 hammer_cache_node(&ip->cache[1], cursor->node);
2001
2002         if (error == EDEADLK) {
2003                 hammer_done_cursor(cursor);
2004                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2005                 if (error == 0)
2006                         goto retry;
2007         }
2008         if (error == ENOENT)
2009                 error = 0;
2010         return(error);
2011 }
2012
2013 /*
2014  * This backend function deletes the specified record on-disk, similar to
2015  * delete_range but for a specific record.  Unlike the exact deletions
2016  * used when deleting a directory entry this function uses an ASOF search
2017  * like delete_range.
2018  *
2019  * This function may be called with ip->obj_asof set for a slave snapshot,
2020  * so don't use it.  We always delete non-historical records only.
2021  */
2022 static int
2023 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2024                       hammer_btree_leaf_elm_t leaf)
2025 {
2026         hammer_transaction_t trans = cursor->trans;
2027         int error;
2028
2029         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2030 retry:
2031         hammer_normalize_cursor(cursor);
2032         cursor->key_beg = leaf->base;
2033         cursor->asof = HAMMER_MAX_TID;
2034         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2035         cursor->flags |= HAMMER_CURSOR_ASOF;
2036         cursor->flags |= HAMMER_CURSOR_BACKEND;
2037         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2038
2039         error = hammer_btree_lookup(cursor);
2040         if (error == 0) {
2041                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2042         }
2043         if (error == EDEADLK) {
2044                 hammer_done_cursor(cursor);
2045                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2046                 if (error == 0)
2047                         goto retry;
2048         }
2049         return(error);
2050 }
2051
2052 /*
2053  * This function deletes remaining auxillary records when an inode is
2054  * being deleted.  This function explicitly does not delete the
2055  * inode record, directory entry, data, or db records.  Those must be
2056  * properly disposed of prior to this call.
2057  */
2058 int
2059 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2060 {
2061         hammer_transaction_t trans = cursor->trans;
2062         hammer_btree_leaf_elm_t leaf;
2063         int error;
2064
2065         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2066 retry:
2067         hammer_normalize_cursor(cursor);
2068         cursor->key_beg.localization = ip->obj_localization +
2069                                        HAMMER_LOCALIZE_MISC;
2070         cursor->key_beg.obj_id = ip->obj_id;
2071         cursor->key_beg.create_tid = 0;
2072         cursor->key_beg.delete_tid = 0;
2073         cursor->key_beg.obj_type = 0;
2074         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2075         cursor->key_beg.key = HAMMER_MIN_KEY;
2076
2077         cursor->key_end = cursor->key_beg;
2078         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2079         cursor->key_end.key = HAMMER_MAX_KEY;
2080
2081         cursor->asof = ip->obj_asof;
2082         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2083         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2084         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2085         cursor->flags |= HAMMER_CURSOR_BACKEND;
2086
2087         error = hammer_ip_first(cursor);
2088
2089         /*
2090          * Iterate through matching records and mark them as deleted.
2091          */
2092         while (error == 0) {
2093                 leaf = cursor->leaf;
2094
2095                 KKASSERT(leaf->base.delete_tid == 0);
2096
2097                 /*
2098                  * Mark the record and B-Tree entry as deleted.  This will
2099                  * also physically delete the B-Tree entry, record, and
2100                  * data if the retention policy dictates.  The function
2101                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2102                  * to retest the new 'current' element.
2103                  *
2104                  * Directory entries (and delete-on-disk directory entries)
2105                  * must be synced and cannot be deleted.
2106                  */
2107                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2108                 ++*countp;
2109                 if (error)
2110                         break;
2111                 error = hammer_ip_next(cursor);
2112         }
2113         if (cursor->node)
2114                 hammer_cache_node(&ip->cache[1], cursor->node);
2115         if (error == EDEADLK) {
2116                 hammer_done_cursor(cursor);
2117                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2118                 if (error == 0)
2119                         goto retry;
2120         }
2121         if (error == ENOENT)
2122                 error = 0;
2123         return(error);
2124 }
2125
2126 /*
2127  * Delete the record at the current cursor.  On success the cursor will
2128  * be positioned appropriately for an iteration but may no longer be at
2129  * a leaf node.
2130  *
2131  * This routine is only called from the backend.
2132  *
2133  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2134  * cursor and retry.
2135  */
2136 int
2137 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2138                         hammer_tid_t tid)
2139 {
2140         hammer_record_t iprec;
2141         hammer_mount_t hmp;
2142         int error;
2143
2144         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2145         KKASSERT(tid != 0);
2146         hmp = cursor->node->hmp;
2147
2148         /*
2149          * In-memory (unsynchronized) records can simply be freed.  This
2150          * only occurs in range iterations since all other records are
2151          * individually synchronized.  Thus there should be no confusion with
2152          * the interlock.
2153          *
2154          * An in-memory record may be deleted before being committed to disk,
2155          * but could have been accessed in the mean time.  The reservation
2156          * code will deal with the case.
2157          */
2158         if (hammer_cursor_inmem(cursor)) {
2159                 iprec = cursor->iprec;
2160                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2161                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2162                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2163                 KKASSERT(iprec->ip == ip);
2164                 ++ip->rec_generation;
2165                 return(0);
2166         }
2167
2168         /*
2169          * On-disk records are marked as deleted by updating their delete_tid.
2170          * This does not effect their position in the B-Tree (which is based
2171          * on their create_tid).
2172          *
2173          * Frontend B-Tree operations track inodes so we tell
2174          * hammer_delete_at_cursor() not to.
2175          */
2176         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2177
2178         if (error == 0) {
2179                 error = hammer_delete_at_cursor(
2180                                 cursor,
2181                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2182                                 cursor->trans->tid,
2183                                 cursor->trans->time32,
2184                                 0, NULL);
2185         }
2186         return(error);
2187 }
2188
2189 /*
2190  * Delete the B-Tree element at the current cursor and do any necessary
2191  * mirror propagation.
2192  *
2193  * The cursor must be properly positioned for an iteration on return but
2194  * may be pointing at an internal element.
2195  *
2196  * An element can be un-deleted by passing a delete_tid of 0 with
2197  * HAMMER_DELETE_ADJUST.
2198  */
2199 int
2200 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2201                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2202                         int track, int64_t *stat_bytes)
2203 {
2204         struct hammer_btree_leaf_elm save_leaf;
2205         hammer_transaction_t trans;
2206         hammer_btree_leaf_elm_t leaf;
2207         hammer_node_t node;
2208         hammer_btree_elm_t elm;
2209         hammer_off_t data_offset;
2210         int32_t data_len;
2211         u_int16_t rec_type;
2212         int error;
2213         int icount;
2214         int doprop;
2215
2216         error = hammer_cursor_upgrade(cursor);
2217         if (error)
2218                 return(error);
2219
2220         trans = cursor->trans;
2221         node = cursor->node;
2222         elm = &node->ondisk->elms[cursor->index];
2223         leaf = &elm->leaf;
2224         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2225
2226         hammer_sync_lock_sh(trans);
2227         doprop = 0;
2228         icount = 0;
2229
2230         /*
2231          * Adjust the delete_tid.  Update the mirror_tid propagation field
2232          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2233          */
2234         if (delete_flags & HAMMER_DELETE_ADJUST) {
2235                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2236                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2237                                 icount = -1;
2238                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2239                                 icount = 1;
2240                 }
2241
2242                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2243                 elm->leaf.base.delete_tid = delete_tid;
2244                 elm->leaf.delete_ts = delete_ts;
2245                 hammer_modify_node_done(node);
2246
2247                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2248                         hammer_modify_node_field(trans, node, mirror_tid);
2249                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2250                         hammer_modify_node_done(node);
2251                         doprop = 1;
2252                         if (hammer_debug_general & 0x0002) {
2253                                 kprintf("delete_at_cursor: propagate %016llx"
2254                                         " @%016llx\n",
2255                                         (long long)elm->leaf.base.delete_tid,
2256                                         (long long)node->node_offset);
2257                         }
2258                 }
2259
2260                 /*
2261                  * Adjust for the iteration.  We have deleted the current
2262                  * element and want to clear ATEDISK so the iteration does
2263                  * not skip the element after, which now becomes the current
2264                  * element.  This element must be re-tested if doing an
2265                  * iteration, which is handled by the RETEST flag.
2266                  */
2267                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2268                         cursor->flags |= HAMMER_CURSOR_RETEST;
2269                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2270                 }
2271
2272                 /*
2273                  * An on-disk record cannot have the same delete_tid
2274                  * as its create_tid.  In a chain of record updates
2275                  * this could result in a duplicate record.
2276                  */
2277                 KKASSERT(elm->leaf.base.delete_tid !=
2278                          elm->leaf.base.create_tid);
2279         }
2280
2281         /*
2282          * Destroy the B-Tree element if asked (typically if a nohistory
2283          * file or mount, or when called by the pruning code).
2284          *
2285          * Adjust the ATEDISK flag to properly support iterations.
2286          */
2287         if (delete_flags & HAMMER_DELETE_DESTROY) {
2288                 data_offset = elm->leaf.data_offset;
2289                 data_len = elm->leaf.data_len;
2290                 rec_type = elm->leaf.base.rec_type;
2291                 if (doprop) {
2292                         save_leaf = elm->leaf;
2293                         leaf = &save_leaf;
2294                 }
2295                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2296                     elm->leaf.base.delete_tid == 0) {
2297                         icount = -1;
2298                 }
2299
2300                 error = hammer_btree_delete(cursor);
2301                 if (error == 0) {
2302                         /*
2303                          * The deletion moves the next element (if any) to
2304                          * the current element position.  We must clear
2305                          * ATEDISK so this element is not skipped and we
2306                          * must set RETEST to force any iteration to re-test
2307                          * the element.
2308                          */
2309                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2310                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2311                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2312                         }
2313                 }
2314                 if (error == 0) {
2315                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2316                         case HAMMER_ZONE_LARGE_DATA:
2317                         case HAMMER_ZONE_SMALL_DATA:
2318                         case HAMMER_ZONE_META:
2319                                 hammer_blockmap_free(trans,
2320                                                      data_offset, data_len);
2321                                 break;
2322                         default:
2323                                 break;
2324                         }
2325                 }
2326         }
2327
2328         /*
2329          * Track inode count and next_tid.  This is used by the mirroring
2330          * and PFS code.  icount can be negative, zero, or positive.
2331          */
2332         if (error == 0 && track) {
2333                 if (icount) {
2334                         hammer_modify_volume_field(trans, trans->rootvol,
2335                                                    vol0_stat_inodes);
2336                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2337                         hammer_modify_volume_done(trans->rootvol);
2338                 }
2339                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2340                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2341                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2342                         hammer_modify_volume_done(trans->rootvol);
2343                 }
2344         }
2345
2346         /*
2347          * mirror_tid propagation occurs if the node's mirror_tid had to be
2348          * updated while adjusting the delete_tid.
2349          *
2350          * This occurs when deleting even in nohistory mode, but does not
2351          * occur when pruning an already-deleted node.
2352          *
2353          * cursor->ip is NULL when called from the pruning, mirroring,
2354          * and pfs code.  If non-NULL propagation will be conditionalized
2355          * on whether the PFS is in no-history mode or not.
2356          */
2357         if (doprop) {
2358                 if (cursor->ip)
2359                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2360                 else
2361                         hammer_btree_do_propagation(cursor, NULL, leaf);
2362         }
2363         hammer_sync_unlock(trans);
2364         return (error);
2365 }
2366
2367 /*
2368  * Determine whether we can remove a directory.  This routine checks whether
2369  * a directory is empty or not and enforces flush connectivity.
2370  *
2371  * Flush connectivity requires that we block if the target directory is
2372  * currently flushing, otherwise it may not end up in the same flush group.
2373  *
2374  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2375  */
2376 int
2377 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2378 {
2379         struct hammer_cursor cursor;
2380         int error;
2381
2382         /*
2383          * Check directory empty
2384          */
2385         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2386
2387         cursor.key_beg.localization = ip->obj_localization +
2388                                       hammer_dir_localization(ip);
2389         cursor.key_beg.obj_id = ip->obj_id;
2390         cursor.key_beg.create_tid = 0;
2391         cursor.key_beg.delete_tid = 0;
2392         cursor.key_beg.obj_type = 0;
2393         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2394         cursor.key_beg.key = HAMMER_MIN_KEY;
2395
2396         cursor.key_end = cursor.key_beg;
2397         cursor.key_end.rec_type = 0xFFFF;
2398         cursor.key_end.key = HAMMER_MAX_KEY;
2399
2400         cursor.asof = ip->obj_asof;
2401         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2402
2403         error = hammer_ip_first(&cursor);
2404         if (error == ENOENT)
2405                 error = 0;
2406         else if (error == 0)
2407                 error = ENOTEMPTY;
2408         hammer_done_cursor(&cursor);
2409         return(error);
2410 }
2411