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_LOCALIZE_MISC;
 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 cluster.  The
1172          * target key should not exist.  If we are creating a directory entry
1173          * we may have to iterate the low 32 bits of the key to find an unused
1174          * 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", 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, &error);
1223                 if (bdata == NULL)
1224                         goto done_unlock;
1225                 hammer_crc_set_leaf(record->data, &record->leaf);
1226                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1227                 bcopy(record->data, bdata, record->leaf.data_len);
1228                 hammer_modify_buffer_done(cursor->data_buffer);
1229         } else {
1230                 /*
1231                  * Wholely cached record, without data.
1232                  */
1233                 record->leaf.data_offset = 0;
1234                 record->leaf.data_crc = 0;
1235         }
1236
1237         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1238         if (hammer_debug_inode && error)
1239                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n", error, cursor->node->node_offset, cursor->index, record->leaf.base.key);
1240
1241         /*
1242          * Our record is on-disk and we normally mark the in-memory version
1243          * as having been committed (and not BE-deleted).
1244          *
1245          * If the record represented a directory deletion but we had to
1246          * sync a valid directory entry to disk due to dependancies,
1247          * we must convert the record to a covering delete so the
1248          * frontend does not have visibility on the synced entry.
1249          */
1250         if (error == 0) {
1251                 if (doprop) {
1252                         hammer_btree_do_propagation(cursor,
1253                                                     record->ip->pfsm,
1254                                                     &record->leaf);
1255                 }
1256                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1257                         /*
1258                          * Must convert deleted directory entry add
1259                          * to a directory entry delete.
1260                          */
1261                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1262                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1263                         record->type = HAMMER_MEM_RECORD_DEL;
1264                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1265                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1266                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1267                                                  HAMMER_RECF_DELETED_BE)) == 0);
1268                         /* converted record is not yet committed */
1269                         /* hammer_flush_record_done takes care of the rest */
1270                 } else {
1271                         /*
1272                          * Everything went fine and we are now done with
1273                          * this record.
1274                          */
1275                         record->flags |= HAMMER_RECF_COMMITTED;
1276                         ++record->ip->rec_generation;
1277                 }
1278         } else {
1279                 if (record->leaf.data_offset) {
1280                         hammer_blockmap_free(trans, record->leaf.data_offset,
1281                                              record->leaf.data_len);
1282                 }
1283         }
1284 done_unlock:
1285         hammer_sync_unlock(trans);
1286 done:
1287         return(error);
1288 }
1289
1290 /*
1291  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1292  * entry's key is used to deal with hash collisions in the upper 32 bits.
1293  * A unique 64 bit key is generated in-memory and may be regenerated a
1294  * second time when the directory record is flushed to the on-disk B-Tree.
1295  *
1296  * A referenced record is passed to this function.  This function
1297  * eats the reference.  If an error occurs the record will be deleted.
1298  *
1299  * A copy of the temporary record->data pointer provided by the caller
1300  * will be made.
1301  */
1302 int
1303 hammer_mem_add(hammer_record_t record)
1304 {
1305         hammer_mount_t hmp = record->ip->hmp;
1306
1307         /*
1308          * Make a private copy of record->data
1309          */
1310         if (record->data)
1311                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1312
1313         /*
1314          * Insert into the RB tree.  A unique key should have already
1315          * been selected if this is a directory entry.
1316          */
1317         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1318                 record->flags |= HAMMER_RECF_DELETED_FE;
1319                 hammer_rel_mem_record(record);
1320                 return (EEXIST);
1321         }
1322         ++hmp->count_newrecords;
1323         ++hmp->rsv_recs;
1324         ++record->ip->rsv_recs;
1325         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1326         record->flags |= HAMMER_RECF_ONRBTREE;
1327         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1328         hammer_rel_mem_record(record);
1329         return(0);
1330 }
1331
1332 /************************************************************************
1333  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1334  ************************************************************************
1335  *
1336  * These functions augment the B-Tree scanning functions in hammer_btree.c
1337  * by merging in-memory records with on-disk records.
1338  */
1339
1340 /*
1341  * Locate a particular record either in-memory or on-disk.
1342  *
1343  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1344  * NOT be called to iterate results.
1345  */
1346 int
1347 hammer_ip_lookup(hammer_cursor_t cursor)
1348 {
1349         int error;
1350
1351         /*
1352          * If the element is in-memory return it without searching the
1353          * on-disk B-Tree
1354          */
1355         KKASSERT(cursor->ip);
1356         error = hammer_mem_lookup(cursor);
1357         if (error == 0) {
1358                 cursor->leaf = &cursor->iprec->leaf;
1359                 return(error);
1360         }
1361         if (error != ENOENT)
1362                 return(error);
1363
1364         /*
1365          * If the inode has on-disk components search the on-disk B-Tree.
1366          */
1367         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1368                 return(error);
1369         error = hammer_btree_lookup(cursor);
1370         if (error == 0)
1371                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1372         return(error);
1373 }
1374
1375 /*
1376  * Helper for hammer_ip_first()/hammer_ip_next()
1377  *
1378  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1379  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1380  * state if hammer_ip_next() needs to re-seek.
1381  */
1382 static __inline
1383 int
1384 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1385 {
1386         hammer_inode_t ip = cursor->ip;
1387         int error;
1388
1389         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1390                 error = hammer_btree_lookup(cursor);
1391                 if (error == ENOENT || error == EDEADLK) {
1392                         if (hammer_debug_general & 0x2000)
1393                                 kprintf("error %d node %p %016llx index %d\n", error, cursor->node, cursor->node->node_offset, cursor->index);
1394                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1395                         error = hammer_btree_iterate(cursor);
1396                 }
1397                 if (error == 0) {
1398                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1399                                            HAMMER_CURSOR_ATEDISK);
1400                 } else {
1401                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1402                                          HAMMER_CURSOR_ATEDISK;
1403                         if (error == ENOENT)
1404                                 error = 0;
1405                 }
1406         } else {
1407                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1408                 error = 0;
1409         }
1410         return(error);
1411 }
1412
1413 /*
1414  * Helper for hammer_ip_next()
1415  *
1416  * The caller has determined that the media cursor is further along than the
1417  * memory cursor and must be reseeked after a generation number change.
1418  */
1419 static
1420 int
1421 _hammer_ip_reseek(hammer_cursor_t cursor)
1422 {
1423         struct hammer_base_elm save;
1424         hammer_btree_elm_t elm;
1425         int error;
1426         int r;
1427         int again = 0;
1428
1429         /*
1430          * Do the re-seek.
1431          */
1432         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1433                 (long long)cursor->ip->obj_id);
1434         save = cursor->key_beg;
1435         cursor->key_beg = cursor->iprec->leaf.base;
1436         error = _hammer_ip_seek_btree(cursor);
1437         KKASSERT(error == 0);
1438         cursor->key_beg = save;
1439
1440         /*
1441          * If the memory record was previous returned to
1442          * the caller and the media record matches
1443          * (-1/+1: only create_tid differs), then iterate
1444          * the media record to avoid a double result.
1445          */
1446         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1447             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1448                 elm = &cursor->node->ondisk->elms[cursor->index];
1449                 r = hammer_btree_cmp(&elm->base,
1450                                      &cursor->iprec->leaf.base);
1451                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1452                         if (r >= -1 && r <= 1) {
1453                                 kprintf("HAMMER: Debug: iterated after "
1454                                         "re-seek (asof r=%d)\n", r);
1455                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1456                                 again = 1;
1457                         }
1458                 } else {
1459                         if (r == 0) {
1460                                 kprintf("HAMMER: Debug: iterated after "
1461                                         "re-seek\n");
1462                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1463                                 again = 1;
1464                         }
1465                 }
1466         }
1467         return(again);
1468 }
1469
1470 /*
1471  * Locate the first record within the cursor's key_beg/key_end range,
1472  * restricted to a particular inode.  0 is returned on success, ENOENT
1473  * if no records matched the requested range, or some other error.
1474  *
1475  * When 0 is returned hammer_ip_next() may be used to iterate additional
1476  * records within the requested range.
1477  *
1478  * This function can return EDEADLK, requiring the caller to terminate
1479  * the cursor and try again.
1480  */
1481
1482 int
1483 hammer_ip_first(hammer_cursor_t cursor)
1484 {
1485         hammer_inode_t ip = cursor->ip;
1486         int error;
1487
1488         KKASSERT(ip != NULL);
1489
1490         /*
1491          * Clean up fields and setup for merged scan
1492          */
1493         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1494
1495         /*
1496          * Search the in-memory record list (Red-Black tree).  Unlike the
1497          * B-Tree search, mem_first checks for records in the range.
1498          *
1499          * This function will setup both ATEMEM and MEMEOF properly for
1500          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1501          */
1502         hammer_mem_first(cursor);
1503
1504         /*
1505          * Detect generation changes during blockages, including
1506          * blockages which occur on the initial btree search.
1507          */
1508         cursor->rec_generation = cursor->ip->rec_generation;
1509
1510         /*
1511          * Initial search and result
1512          */
1513         error = _hammer_ip_seek_btree(cursor);
1514         if (error == 0)
1515                 error = hammer_ip_next(cursor);
1516
1517         return (error);
1518 }
1519
1520 /*
1521  * Retrieve the next record in a merged iteration within the bounds of the
1522  * cursor.  This call may be made multiple times after the cursor has been
1523  * initially searched with hammer_ip_first().
1524  *
1525  * There are numerous special cases in this code to deal with races between
1526  * in-memory records and on-media records.
1527  *
1528  * 0 is returned on success, ENOENT if no further records match the
1529  * requested range, or some other error code is returned.
1530  */
1531 int
1532 hammer_ip_next(hammer_cursor_t cursor)
1533 {
1534         hammer_btree_elm_t elm;
1535         hammer_record_t rec;
1536         hammer_record_t tmprec;
1537         int error;
1538         int r;
1539
1540 again:
1541         /*
1542          * Get the next on-disk record
1543          *
1544          * NOTE: If we deleted the last on-disk record we had scanned
1545          *       ATEDISK will be clear and RETEST will be set, forcing
1546          *       a call to iterate.  The fact that ATEDISK is clear causes
1547          *       iterate to re-test the 'current' element.  If ATEDISK is
1548          *       set, iterate will skip the 'current' element.
1549          */
1550         error = 0;
1551         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1552                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1553                                      HAMMER_CURSOR_RETEST)) {
1554                         error = hammer_btree_iterate(cursor);
1555                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1556                         if (error == 0) {
1557                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1558                                 hammer_cache_node(&cursor->ip->cache[1],
1559                                                   cursor->node);
1560                         } else if (error == ENOENT) {
1561                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1562                                                  HAMMER_CURSOR_ATEDISK;
1563                                 error = 0;
1564                         }
1565                 }
1566         }
1567
1568         /*
1569          * If the generation changed the backend has deleted or committed
1570          * one or more memory records since our last check.
1571          *
1572          * When this case occurs if the disk cursor is > current memory record
1573          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1574          * Since the cursor is ahead it must have not yet been eaten (if
1575          * not at eof anyway). (XXX data offset case?)
1576          *
1577          * NOTE: we are not doing a full check here.  That will be handled
1578          * later on.
1579          *
1580          * If we have exhausted all memory records we do not have to do any
1581          * further seeks.
1582          */
1583         while (cursor->rec_generation != cursor->ip->rec_generation &&
1584                error == 0
1585         ) {
1586                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1587                 cursor->rec_generation = cursor->ip->rec_generation;
1588                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1589                         break;
1590                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1591                         r = 1;
1592                 } else {
1593                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1594                         elm = &cursor->node->ondisk->elms[cursor->index];
1595                         r = hammer_btree_cmp(&elm->base,
1596                                              &cursor->iprec->leaf.base);
1597                 }
1598
1599                 /*
1600                  * Do we re-seek the media cursor?
1601                  */
1602                 if (r > 0) {
1603                         if (_hammer_ip_reseek(cursor))
1604                                 goto again;
1605                 }
1606         }
1607
1608         /*
1609          * We can now safely get the next in-memory record.  We cannot
1610          * block here.
1611          *
1612          * hammer_rec_scan_cmp:  Is the record still in our general range,
1613          *                       (non-inclusive of snapshot exclusions)?
1614          * hammer_rec_scan_callback: Is the record in our snapshot?
1615          */
1616         tmprec = NULL;
1617         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1618                 /*
1619                  * If the current memory record was eaten then get the next
1620                  * one.  Stale records are skipped.
1621                  */
1622                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1623                         tmprec = cursor->iprec;
1624                         cursor->iprec = NULL;
1625                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1626                         while (rec) {
1627                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1628                                         break;
1629                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1630                                         break;
1631                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1632                         }
1633                         if (cursor->iprec) {
1634                                 KKASSERT(cursor->iprec == rec);
1635                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1636                         } else {
1637                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1638                         }
1639                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1640                 }
1641         }
1642
1643         /*
1644          * MEMORY RECORD VALIDITY TEST
1645          *
1646          * (We still can't block, which is why tmprec is being held so
1647          * long).
1648          *
1649          * If the memory record is no longer valid we skip it.  It may
1650          * have been deleted by the frontend.  If it was deleted or
1651          * committed by the backend the generation change re-seeked the
1652          * disk cursor and the record will be present there.
1653          */
1654         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1655                 KKASSERT(cursor->iprec);
1656                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1657                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1658                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1659                         if (tmprec)
1660                                 hammer_rel_mem_record(tmprec);
1661                         goto again;
1662                 }
1663         }
1664         if (tmprec)
1665                 hammer_rel_mem_record(tmprec);
1666
1667         /*
1668          * Extract either the disk or memory record depending on their
1669          * relative position.
1670          */
1671         error = 0;
1672         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1673         case 0:
1674                 /*
1675                  * Both entries valid.   Compare the entries and nominally
1676                  * return the first one in the sort order.  Numerous cases
1677                  * require special attention, however.
1678                  */
1679                 elm = &cursor->node->ondisk->elms[cursor->index];
1680                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1681
1682                 /*
1683                  * If the two entries differ only by their key (-2/2) or
1684                  * create_tid (-1/1), and are DATA records, we may have a
1685                  * nominal match.  We have to calculate the base file
1686                  * offset of the data.
1687                  */
1688                 if (r <= 2 && r >= -2 && r != 0 &&
1689                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1690                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1691                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1692                         int64_t base2 = cursor->iprec->leaf.base.key -
1693                                         cursor->iprec->leaf.data_len;
1694                         if (base1 == base2)
1695                                 r = 0;
1696                 }
1697
1698                 if (r < 0) {
1699                         error = hammer_btree_extract(cursor,
1700                                                      HAMMER_CURSOR_GET_LEAF);
1701                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1702                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1703                         break;
1704                 }
1705
1706                 /*
1707                  * If the entries match exactly the memory entry is either
1708                  * an on-disk directory entry deletion or a bulk data
1709                  * overwrite.  If it is a directory entry deletion we eat
1710                  * both entries.
1711                  *
1712                  * For the bulk-data overwrite case it is possible to have
1713                  * visibility into both, which simply means the syncer
1714                  * hasn't gotten around to doing the delete+insert sequence
1715                  * on the B-Tree.  Use the memory entry and throw away the
1716                  * on-disk entry.
1717                  *
1718                  * If the in-memory record is not either of these we
1719                  * probably caught the syncer while it was syncing it to
1720                  * the media.  Since we hold a shared lock on the cursor,
1721                  * the in-memory record had better be marked deleted at
1722                  * this point.
1723                  */
1724                 if (r == 0) {
1725                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1726                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1727                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1728                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1729                                         goto again;
1730                                 }
1731                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1732                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1733                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1734                                 }
1735                                 /* fall through to memory entry */
1736                         } else {
1737                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1738                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1739                                 goto again;
1740                         }
1741                 }
1742                 /* fall through to the memory entry */
1743         case HAMMER_CURSOR_ATEDISK:
1744                 /*
1745                  * Only the memory entry is valid.
1746                  */
1747                 cursor->leaf = &cursor->iprec->leaf;
1748                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1749                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1750
1751                 /*
1752                  * If the memory entry is an on-disk deletion we should have
1753                  * also had found a B-Tree record.  If the backend beat us
1754                  * to it it would have interlocked the cursor and we should
1755                  * have seen the in-memory record marked DELETED_FE.
1756                  */
1757                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1758                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1759                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1760                 }
1761                 break;
1762         case HAMMER_CURSOR_ATEMEM:
1763                 /*
1764                  * Only the disk entry is valid
1765                  */
1766                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1767                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1768                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1769                 break;
1770         default:
1771                 /*
1772                  * Neither entry is valid
1773                  *
1774                  * XXX error not set properly
1775                  */
1776                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1777                 cursor->leaf = NULL;
1778                 error = ENOENT;
1779                 break;
1780         }
1781         return(error);
1782 }
1783
1784 /*
1785  * Resolve the cursor->data pointer for the current cursor position in
1786  * a merged iteration.
1787  */
1788 int
1789 hammer_ip_resolve_data(hammer_cursor_t cursor)
1790 {
1791         hammer_record_t record;
1792         int error;
1793
1794         if (hammer_cursor_inmem(cursor)) {
1795                 /*
1796                  * The data associated with an in-memory record is usually
1797                  * kmalloced, but reserve-ahead data records will have an
1798                  * on-disk reference.
1799                  *
1800                  * NOTE: Reserve-ahead data records must be handled in the
1801                  * context of the related high level buffer cache buffer
1802                  * to interlock against async writes.
1803                  */
1804                 record = cursor->iprec;
1805                 cursor->data = record->data;
1806                 error = 0;
1807                 if (cursor->data == NULL) {
1808                         KKASSERT(record->leaf.base.rec_type ==
1809                                  HAMMER_RECTYPE_DATA);
1810                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1811                                                     record->leaf.data_offset,
1812                                                     record->leaf.data_len,
1813                                                     &error,
1814                                                     &cursor->data_buffer);
1815                 }
1816         } else {
1817                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1818                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1819         }
1820         return(error);
1821 }
1822
1823 /*
1824  * Backend truncation / record replacement - delete records in range.
1825  *
1826  * Delete all records within the specified range for inode ip.  In-memory
1827  * records still associated with the frontend are ignored.
1828  *
1829  * If truncating is non-zero in-memory records associated with the back-end
1830  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1831  *
1832  * NOTES:
1833  *
1834  *      * An unaligned range will cause new records to be added to cover
1835  *        the edge cases. (XXX not implemented yet).
1836  *
1837  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1838  *        also do not deal with unaligned ranges.
1839  *
1840  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1841  *
1842  *      * Record keys for regular file data have to be special-cased since
1843  *        they indicate the end of the range (key = base + bytes).
1844  *
1845  *      * This function may be asked to delete ridiculously huge ranges, for
1846  *        example if someone truncates or removes a 1TB regular file.  We
1847  *        must be very careful on restarts and we may have to stop w/
1848  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1849  */
1850 int
1851 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1852                        int64_t ran_beg, int64_t ran_end, int truncating)
1853 {
1854         hammer_transaction_t trans = cursor->trans;
1855         hammer_btree_leaf_elm_t leaf;
1856         int error;
1857         int64_t off;
1858         int64_t tmp64;
1859
1860 #if 0
1861         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1862 #endif
1863
1864         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1865 retry:
1866         hammer_normalize_cursor(cursor);
1867         cursor->key_beg.localization = ip->obj_localization +
1868                                        HAMMER_LOCALIZE_MISC;
1869         cursor->key_beg.obj_id = ip->obj_id;
1870         cursor->key_beg.create_tid = 0;
1871         cursor->key_beg.delete_tid = 0;
1872         cursor->key_beg.obj_type = 0;
1873
1874         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1875                 cursor->key_beg.key = ran_beg;
1876                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1877         } else {
1878                 /*
1879                  * The key in the B-Tree is (base+bytes), so the first possible
1880                  * matching key is ran_beg + 1.
1881                  */
1882                 cursor->key_beg.key = ran_beg + 1;
1883                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1884         }
1885
1886         cursor->key_end = cursor->key_beg;
1887         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1888                 cursor->key_end.key = ran_end;
1889         } else {
1890                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1891                 if (tmp64 < ran_end)
1892                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1893                 else
1894                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1895         }
1896
1897         cursor->asof = ip->obj_asof;
1898         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1899         cursor->flags |= HAMMER_CURSOR_ASOF;
1900         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1901         cursor->flags |= HAMMER_CURSOR_BACKEND;
1902         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1903
1904         error = hammer_ip_first(cursor);
1905
1906         /*
1907          * Iterate through matching records and mark them as deleted.
1908          */
1909         while (error == 0) {
1910                 leaf = cursor->leaf;
1911
1912                 KKASSERT(leaf->base.delete_tid == 0);
1913                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1914
1915                 /*
1916                  * There may be overlap cases for regular file data.  Also
1917                  * remember the key for a regular file record is (base + len),
1918                  * NOT (base).
1919                  *
1920                  * Note that do to duplicates (mem & media) allowed by
1921                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1922                  */
1923                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1924                         off = leaf->base.key - leaf->data_len;
1925                         /*
1926                          * Check the left edge case.  We currently do not
1927                          * split existing records.
1928                          */
1929                         if (off < ran_beg && leaf->base.key > ran_beg) {
1930                                 panic("hammer left edge case %016llx %d\n",
1931                                         leaf->base.key, leaf->data_len);
1932                         }
1933
1934                         /*
1935                          * Check the right edge case.  Note that the
1936                          * record can be completely out of bounds, which
1937                          * terminates the search.
1938                          *
1939                          * base->key is exclusive of the right edge while
1940                          * ran_end is inclusive of the right edge.  The
1941                          * (key - data_len) left boundary is inclusive.
1942                          *
1943                          * XXX theory-check this test at some point, are
1944                          * we missing a + 1 somewhere?  Note that ran_end
1945                          * could overflow.
1946                          */
1947                         if (leaf->base.key - 1 > ran_end) {
1948                                 if (leaf->base.key - leaf->data_len > ran_end)
1949                                         break;
1950                                 panic("hammer right edge case\n");
1951                         }
1952                 } else {
1953                         off = leaf->base.key;
1954                 }
1955
1956                 /*
1957                  * Delete the record.  When truncating we do not delete
1958                  * in-memory (data) records because they represent data
1959                  * written after the truncation.
1960                  *
1961                  * This will also physically destroy the B-Tree entry and
1962                  * data if the retention policy dictates.  The function
1963                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
1964                  * to retest the new 'current' element.
1965                  */
1966                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1967                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1968                         /*
1969                          * If we have built up too many meta-buffers we risk
1970                          * deadlocking the kernel and must stop.  This can
1971                          * occur when deleting ridiculously huge files.
1972                          * sync_trunc_off is updated so the next cycle does
1973                          * not re-iterate records we have already deleted.
1974                          *
1975                          * This is only done with formal truncations.
1976                          */
1977                         if (truncating > 1 && error == 0 &&
1978                             hammer_flusher_meta_limit(ip->hmp)) {
1979                                 ip->sync_trunc_off = off;
1980                                 error = EWOULDBLOCK;
1981                         }
1982                 }
1983                 if (error)
1984                         break;
1985                 ran_beg = off;  /* for restart */
1986                 error = hammer_ip_next(cursor);
1987         }
1988         if (cursor->node)
1989                 hammer_cache_node(&ip->cache[1], cursor->node);
1990
1991         if (error == EDEADLK) {
1992                 hammer_done_cursor(cursor);
1993                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
1994                 if (error == 0)
1995                         goto retry;
1996         }
1997         if (error == ENOENT)
1998                 error = 0;
1999         return(error);
2000 }
2001
2002 /*
2003  * This backend function deletes the specified record on-disk, similar to
2004  * delete_range but for a specific record.  Unlike the exact deletions
2005  * used when deleting a directory entry this function uses an ASOF search
2006  * like delete_range.
2007  *
2008  * This function may be called with ip->obj_asof set for a slave snapshot,
2009  * so don't use it.  We always delete non-historical records only.
2010  */
2011 static int
2012 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2013                       hammer_btree_leaf_elm_t leaf)
2014 {
2015         hammer_transaction_t trans = cursor->trans;
2016         int error;
2017
2018         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2019 retry:
2020         hammer_normalize_cursor(cursor);
2021         cursor->key_beg = leaf->base;
2022         cursor->asof = HAMMER_MAX_TID;
2023         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2024         cursor->flags |= HAMMER_CURSOR_ASOF;
2025         cursor->flags |= HAMMER_CURSOR_BACKEND;
2026         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2027
2028         error = hammer_btree_lookup(cursor);
2029         if (error == 0) {
2030                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2031         }
2032         if (error == EDEADLK) {
2033                 hammer_done_cursor(cursor);
2034                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2035                 if (error == 0)
2036                         goto retry;
2037         }
2038         return(error);
2039 }
2040
2041 /*
2042  * This function deletes remaining auxillary records when an inode is
2043  * being deleted.  This function explicitly does not delete the
2044  * inode record, directory entry, data, or db records.  Those must be
2045  * properly disposed of prior to this call.
2046  */
2047 int
2048 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2049 {
2050         hammer_transaction_t trans = cursor->trans;
2051         hammer_btree_leaf_elm_t leaf;
2052         int error;
2053
2054         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2055 retry:
2056         hammer_normalize_cursor(cursor);
2057         cursor->key_beg.localization = ip->obj_localization +
2058                                        HAMMER_LOCALIZE_MISC;
2059         cursor->key_beg.obj_id = ip->obj_id;
2060         cursor->key_beg.create_tid = 0;
2061         cursor->key_beg.delete_tid = 0;
2062         cursor->key_beg.obj_type = 0;
2063         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2064         cursor->key_beg.key = HAMMER_MIN_KEY;
2065
2066         cursor->key_end = cursor->key_beg;
2067         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2068         cursor->key_end.key = HAMMER_MAX_KEY;
2069
2070         cursor->asof = ip->obj_asof;
2071         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2072         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2073         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2074         cursor->flags |= HAMMER_CURSOR_BACKEND;
2075
2076         error = hammer_ip_first(cursor);
2077
2078         /*
2079          * Iterate through matching records and mark them as deleted.
2080          */
2081         while (error == 0) {
2082                 leaf = cursor->leaf;
2083
2084                 KKASSERT(leaf->base.delete_tid == 0);
2085
2086                 /*
2087                  * Mark the record and B-Tree entry as deleted.  This will
2088                  * also physically delete the B-Tree entry, record, and
2089                  * data if the retention policy dictates.  The function
2090                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2091                  * to retest the new 'current' element.
2092                  *
2093                  * Directory entries (and delete-on-disk directory entries)
2094                  * must be synced and cannot be deleted.
2095                  */
2096                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2097                 ++*countp;
2098                 if (error)
2099                         break;
2100                 error = hammer_ip_next(cursor);
2101         }
2102         if (cursor->node)
2103                 hammer_cache_node(&ip->cache[1], cursor->node);
2104         if (error == EDEADLK) {
2105                 hammer_done_cursor(cursor);
2106                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2107                 if (error == 0)
2108                         goto retry;
2109         }
2110         if (error == ENOENT)
2111                 error = 0;
2112         return(error);
2113 }
2114
2115 /*
2116  * Delete the record at the current cursor.  On success the cursor will
2117  * be positioned appropriately for an iteration but may no longer be at
2118  * a leaf node.
2119  *
2120  * This routine is only called from the backend.
2121  *
2122  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2123  * cursor and retry.
2124  */
2125 int
2126 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2127                         hammer_tid_t tid)
2128 {
2129         hammer_record_t iprec;
2130         hammer_mount_t hmp;
2131         int error;
2132
2133         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2134         KKASSERT(tid != 0);
2135         hmp = cursor->node->hmp;
2136
2137         /*
2138          * In-memory (unsynchronized) records can simply be freed.  This
2139          * only occurs in range iterations since all other records are
2140          * individually synchronized.  Thus there should be no confusion with
2141          * the interlock.
2142          *
2143          * An in-memory record may be deleted before being committed to disk,
2144          * but could have been accessed in the mean time.  The reservation
2145          * code will deal with the case.
2146          */
2147         if (hammer_cursor_inmem(cursor)) {
2148                 iprec = cursor->iprec;
2149                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2150                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2151                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2152                 KKASSERT(iprec->ip == ip);
2153                 ++ip->rec_generation;
2154                 return(0);
2155         }
2156
2157         /*
2158          * On-disk records are marked as deleted by updating their delete_tid.
2159          * This does not effect their position in the B-Tree (which is based
2160          * on their create_tid).
2161          *
2162          * Frontend B-Tree operations track inodes so we tell
2163          * hammer_delete_at_cursor() not to.
2164          */
2165         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2166
2167         if (error == 0) {
2168                 error = hammer_delete_at_cursor(
2169                                 cursor,
2170                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2171                                 cursor->trans->tid,
2172                                 cursor->trans->time32,
2173                                 0, NULL);
2174         }
2175         return(error);
2176 }
2177
2178 /*
2179  * Delete the B-Tree element at the current cursor and do any necessary
2180  * mirror propagation.
2181  *
2182  * The cursor must be properly positioned for an iteration on return but
2183  * may be pointing at an internal element.
2184  *
2185  * An element can be un-deleted by passing a delete_tid of 0 with
2186  * HAMMER_DELETE_ADJUST.
2187  */
2188 int
2189 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2190                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2191                         int track, int64_t *stat_bytes)
2192 {
2193         struct hammer_btree_leaf_elm save_leaf;
2194         hammer_transaction_t trans;
2195         hammer_btree_leaf_elm_t leaf;
2196         hammer_node_t node;
2197         hammer_btree_elm_t elm;
2198         hammer_off_t data_offset;
2199         int32_t data_len;
2200         u_int16_t rec_type;
2201         int error;
2202         int icount;
2203         int doprop;
2204
2205         error = hammer_cursor_upgrade(cursor);
2206         if (error)
2207                 return(error);
2208
2209         trans = cursor->trans;
2210         node = cursor->node;
2211         elm = &node->ondisk->elms[cursor->index];
2212         leaf = &elm->leaf;
2213         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2214
2215         hammer_sync_lock_sh(trans);
2216         doprop = 0;
2217         icount = 0;
2218
2219         /*
2220          * Adjust the delete_tid.  Update the mirror_tid propagation field
2221          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2222          */
2223         if (delete_flags & HAMMER_DELETE_ADJUST) {
2224                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2225                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2226                                 icount = -1;
2227                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2228                                 icount = 1;
2229                 }
2230
2231                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2232                 elm->leaf.base.delete_tid = delete_tid;
2233                 elm->leaf.delete_ts = delete_ts;
2234                 hammer_modify_node_done(node);
2235
2236                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2237                         hammer_modify_node_field(trans, node, mirror_tid);
2238                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2239                         hammer_modify_node_done(node);
2240                         doprop = 1;
2241                         if (hammer_debug_general & 0x0002) {
2242                                 kprintf("delete_at_cursor: propagate %016llx"
2243                                         " @%016llx\n",
2244                                         elm->leaf.base.delete_tid,
2245                                         node->node_offset);
2246                         }
2247                 }
2248
2249                 /*
2250                  * Adjust for the iteration.  We have deleted the current
2251                  * element and want to clear ATEDISK so the iteration does
2252                  * not skip the element after, which now becomes the current
2253                  * element.  This element must be re-tested if doing an
2254                  * iteration, which is handled by the RETEST flag.
2255                  */
2256                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2257                         cursor->flags |= HAMMER_CURSOR_RETEST;
2258                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2259                 }
2260
2261                 /*
2262                  * An on-disk record cannot have the same delete_tid
2263                  * as its create_tid.  In a chain of record updates
2264                  * this could result in a duplicate record.
2265                  */
2266                 KKASSERT(elm->leaf.base.delete_tid !=
2267                          elm->leaf.base.create_tid);
2268         }
2269
2270         /*
2271          * Destroy the B-Tree element if asked (typically if a nohistory
2272          * file or mount, or when called by the pruning code).
2273          *
2274          * Adjust the ATEDISK flag to properly support iterations.
2275          */
2276         if (delete_flags & HAMMER_DELETE_DESTROY) {
2277                 data_offset = elm->leaf.data_offset;
2278                 data_len = elm->leaf.data_len;
2279                 rec_type = elm->leaf.base.rec_type;
2280                 if (doprop) {
2281                         save_leaf = elm->leaf;
2282                         leaf = &save_leaf;
2283                 }
2284                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2285                     elm->leaf.base.delete_tid == 0) {
2286                         icount = -1;
2287                 }
2288
2289                 error = hammer_btree_delete(cursor);
2290                 if (error == 0) {
2291                         /*
2292                          * The deletion moves the next element (if any) to
2293                          * the current element position.  We must clear
2294                          * ATEDISK so this element is not skipped and we
2295                          * must set RETEST to force any iteration to re-test
2296                          * the element.
2297                          */
2298                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2299                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2300                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2301                         }
2302                 }
2303                 if (error == 0) {
2304                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2305                         case HAMMER_ZONE_LARGE_DATA:
2306                         case HAMMER_ZONE_SMALL_DATA:
2307                         case HAMMER_ZONE_META:
2308                                 hammer_blockmap_free(trans,
2309                                                      data_offset, data_len);
2310                                 break;
2311                         default:
2312                                 break;
2313                         }
2314                 }
2315         }
2316
2317         /*
2318          * Track inode count and next_tid.  This is used by the mirroring
2319          * and PFS code.  icount can be negative, zero, or positive.
2320          */
2321         if (error == 0 && track) {
2322                 if (icount) {
2323                         hammer_modify_volume_field(trans, trans->rootvol,
2324                                                    vol0_stat_inodes);
2325                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2326                         hammer_modify_volume_done(trans->rootvol);
2327                 }
2328                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2329                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2330                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2331                         hammer_modify_volume_done(trans->rootvol);
2332                 }
2333         }
2334
2335         /*
2336          * mirror_tid propagation occurs if the node's mirror_tid had to be
2337          * updated while adjusting the delete_tid.
2338          *
2339          * This occurs when deleting even in nohistory mode, but does not
2340          * occur when pruning an already-deleted node.
2341          *
2342          * cursor->ip is NULL when called from the pruning, mirroring,
2343          * and pfs code.  If non-NULL propagation will be conditionalized
2344          * on whether the PFS is in no-history mode or not.
2345          */
2346         if (doprop) {
2347                 if (cursor->ip)
2348                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2349                 else
2350                         hammer_btree_do_propagation(cursor, NULL, leaf);
2351         }
2352         hammer_sync_unlock(trans);
2353         return (error);
2354 }
2355
2356 /*
2357  * Determine whether we can remove a directory.  This routine checks whether
2358  * a directory is empty or not and enforces flush connectivity.
2359  *
2360  * Flush connectivity requires that we block if the target directory is
2361  * currently flushing, otherwise it may not end up in the same flush group.
2362  *
2363  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2364  */
2365 int
2366 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2367 {
2368         struct hammer_cursor cursor;
2369         int error;
2370
2371         /*
2372          * Check directory empty
2373          */
2374         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2375
2376         cursor.key_beg.localization = ip->obj_localization +
2377                                       HAMMER_LOCALIZE_MISC;
2378         cursor.key_beg.obj_id = ip->obj_id;
2379         cursor.key_beg.create_tid = 0;
2380         cursor.key_beg.delete_tid = 0;
2381         cursor.key_beg.obj_type = 0;
2382         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2383         cursor.key_beg.key = HAMMER_MIN_KEY;
2384
2385         cursor.key_end = cursor.key_beg;
2386         cursor.key_end.rec_type = 0xFFFF;
2387         cursor.key_end.key = HAMMER_MAX_KEY;
2388
2389         cursor.asof = ip->obj_asof;
2390         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2391
2392         error = hammer_ip_first(&cursor);
2393         if (error == ENOENT)
2394                 error = 0;
2395         else if (error == 0)
2396                 error = ENOTEMPTY;
2397         hammer_done_cursor(&cursor);
2398         return(error);
2399 }
2400