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         int diddrop;
 367
 368         hammer_unref(&record->lock);
 369
 370         if (record->lock.refs == 0) {
 371                 /*
 372                  * Upon release of the last reference wakeup any waiters.
 373                  * The record structure may get destroyed so callers will
 374                  * loop up and do a relookup.
 375                  *
 376                  * WARNING!  Record must be removed from RB-TREE before we
 377                  * might possibly block.  hammer_test_inode() can block!
 378                  */
 379                 ip = record->ip;
 380                 hmp = ip->hmp;
 381
 382                 /*
 383                  * Upon release of the last reference a record marked deleted
 384                  * by the front or backend, or committed by the backend,
 385                  * is destroyed.
 386                  */
 387                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 388                                      HAMMER_RECF_DELETED_BE |
 389                                      HAMMER_RECF_COMMITTED)) {
 390                         KKASSERT(ip->lock.refs > 0);
 391                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 392
 393                         /*
 394                          * target_ip may have zero refs, we have to ref it
 395                          * to prevent it from being ripped out from under
 396                          * us.
 397                          */
 398                         if ((target_ip = record->target_ip) != NULL) {
 399                                 TAILQ_REMOVE(&target_ip->target_list,
 400                                              record, target_entry);
 401                                 record->target_ip = NULL;
 402                                 hammer_ref(&target_ip->lock);
 403                         }
 404
 405                         /*
 406                          * Remove the record from the B-Tree
 407                          */
 408                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 409                                 RB_REMOVE(hammer_rec_rb_tree,
 410                                           &record->ip->rec_tree,
 411                                           record);
 412                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 413                                 KKASSERT(ip->rsv_recs > 0);
 414                                 diddrop = 1;
 415                         } else {
 416                                 diddrop = 0;
 417                         }
 418
 419                         /*
 420                          * We must wait for any direct-IO to complete before
 421                          * we can destroy the record because the bio may
 422                          * have a reference to it.
 423                          */
 424                         if (record->flags &
 425                            (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL)) {
 426                                 hammer_io_direct_wait(record);
 427                         }
 428
 429                         /*
 430                          * Account for the completion after the direct IO
 431                          * has completed.
 432                          */
 433                         if (diddrop) {
 434                                 --hmp->rsv_recs;
 435                                 --ip->rsv_recs;
 436                                 hmp->rsv_databytes -= record->leaf.data_len;
 437
 438                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 439                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 440                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 441                                         hammer_test_inode(record->ip);
 442                                 }
 443                                 if (ip->rsv_recs == hammer_limit_inode_recs - 1)
 444                                         wakeup(&ip->rsv_recs);
 445                         }
 446
 447                         /*
 448                          * Do this test after removing record from the B-Tree.
 449                          */
 450                         if (target_ip) {
 451                                 hammer_test_inode(target_ip);
 452                                 hammer_rel_inode(target_ip, 0);
 453                         }
 454
 455                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 456                                 --hammer_count_record_datas;
 457                                 kfree(record->data, hmp->m_misc);
 458                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 459                         }
 460
 461                         /*
 462                          * Release the reservation.
 463                          *
 464                          * If the record was not committed we can theoretically
 465                          * undo the reservation.  However, doing so might
 466                          * create weird edge cases with the ordering of
 467                          * direct writes because the related buffer cache
 468                          * elements are per-vnode.  So we don't try.
 469                          */
 470                         if ((resv = record->resv) != NULL) {
 471                                 /* XXX undo leaf.data_offset,leaf.data_len */
 472                                 hammer_blockmap_reserve_complete(hmp, resv);
 473                                 record->resv = NULL;
 474                         }
 475                         record->data = NULL;
 476                         --hammer_count_records;
 477                         kfree(record, hmp->m_misc);
 478                 }
 479         }
 480 }
 481
 482 /*
 483  * Record visibility depends on whether the record is being accessed by
 484  * the backend or the frontend.  Backend tests ignore the frontend delete
 485  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 486  * must also check for commit races.
 487  *
 488  * Return non-zero if the record is visible, zero if it isn't or if it is
 489  * deleted.  Returns 0 if the record has been comitted (unless the special
 490  * delete-visibility flag is set).  A committed record must be located
 491  * via the media B-Tree.  Returns non-zero if the record is good.
 492  *
 493  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 494  * records to be returned.  This is so pending deletions are detected
 495  * when using an iterator to locate an unused hash key, or when we need
 496  * to locate historical records on-disk to destroy.
 497  */
 498 static __inline
 499 int
 500 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 501 {
 502         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 503                 return(1);
 504         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 505                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 506                                      HAMMER_RECF_COMMITTED)) {
 507                         return(0);
 508                 }
 509         } else {
 510                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 511                                      HAMMER_RECF_DELETED_BE |
 512                                      HAMMER_RECF_COMMITTED)) {
 513                         return(0);
 514                 }
 515         }
 516         return(1);
 517 }
 518
 519 /*
 520  * This callback is used as part of the RB_SCAN function for in-memory
 521  * records.  We terminate it (return -1) as soon as we get a match.
 522  *
 523  * This routine is used by frontend code.
 524  *
 525  * The primary compare code does not account for ASOF lookups.  This
 526  * code handles that case as well as a few others.
 527  */
 528 static
 529 int
 530 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 531 {
 532         hammer_cursor_t cursor = data;
 533
 534         /*
 535          * We terminate on success, so this should be NULL on entry.
 536          */
 537         KKASSERT(cursor->iprec == NULL);
 538
 539         /*
 540          * Skip if the record was marked deleted or committed.
 541          */
 542         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 543                 return(0);
 544
 545         /*
 546          * Skip if not visible due to our as-of TID
 547          */
 548         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 549                 if (cursor->asof < rec->leaf.base.create_tid)
 550                         return(0);
 551                 if (rec->leaf.base.delete_tid &&
 552                     cursor->asof >= rec->leaf.base.delete_tid) {
 553                         return(0);
 554                 }
 555         }
 556
 557         /*
 558          * ref the record.  The record is protected from backend B-Tree
 559          * interactions by virtue of the cursor's IP lock.
 560          */
 561         hammer_ref(&rec->lock);
 562
 563         /*
 564          * The record may have been deleted or committed while we
 565          * were blocked.  XXX remove?
 566          */
 567         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 568                 hammer_rel_mem_record(rec);
 569                 return(0);
 570         }
 571
 572         /*
 573          * Set the matching record and stop the scan.
 574          */
 575         cursor->iprec = rec;
 576         return(-1);
 577 }
 578
 579
 580 /*
 581  * Lookup an in-memory record given the key specified in the cursor.  Works
 582  * just like hammer_btree_lookup() but operates on an inode's in-memory
 583  * record list.
 584  *
 585  * The lookup must fail if the record is marked for deferred deletion.
 586  *
 587  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 588  */
 589 static
 590 int
 591 hammer_mem_lookup(hammer_cursor_t cursor)
 592 {
 593         KKASSERT(cursor->ip);
 594         if (cursor->iprec) {
 595                 hammer_rel_mem_record(cursor->iprec);
 596                 cursor->iprec = NULL;
 597         }
 598         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 599                                    hammer_rec_scan_callback, cursor);
 600
 601         return (cursor->iprec ? 0 : ENOENT);
 602 }
 603
 604 /*
 605  * hammer_mem_first() - locate the first in-memory record matching the
 606  * cursor within the bounds of the key range.
 607  *
 608  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 609  * will set ATEMEM the same as MEMEOF, and does not return any error.
 610  */
 611 static
 612 void
 613 hammer_mem_first(hammer_cursor_t cursor)
 614 {
 615         hammer_inode_t ip;
 616
 617         ip = cursor->ip;
 618         KKASSERT(ip != NULL);
 619
 620         if (cursor->iprec) {
 621                 hammer_rel_mem_record(cursor->iprec);
 622                 cursor->iprec = NULL;
 623         }
 624         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 625                                    hammer_rec_scan_callback, cursor);
 626
 627         if (cursor->iprec)
 628                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
 629         else
 630                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
 631 }
 632
 633 /************************************************************************
 634  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 635  ************************************************************************
 636  *
 637  * These functions manipulate in-memory records.  Such records typically
 638  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 639  */
 640
 641 /*
 642  * Add a directory entry (dip,ncp) which references inode (ip).
 643  *
 644  * Note that the low 32 bits of the namekey are set temporarily to create
 645  * a unique in-memory record, and may be modified a second time when the
 646  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 647  * all 0's when synching to disk, which is not handled here.
 648  *
 649  * NOTE: bytes does not include any terminating \0 on name, and name might
 650  * not be terminated.
 651  */
 652 int
 653 hammer_ip_add_directory(struct hammer_transaction *trans,
 654                      struct hammer_inode *dip, const char *name, int bytes,
 655                      struct hammer_inode *ip)
 656 {
 657         struct hammer_cursor cursor;
 658         hammer_record_t record;
 659         int error;
 660         u_int32_t max_iterations;
 661
 662         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 663
 664         record->type = HAMMER_MEM_RECORD_ADD;
 665         record->leaf.base.localization = dip->obj_localization +
 666                                          hammer_dir_localization(dip);
 667         record->leaf.base.obj_id = dip->obj_id;
 668         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 669                                                          &max_iterations);
 670         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 671         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 672         record->data->entry.obj_id = ip->obj_id;
 673         record->data->entry.localization = ip->obj_localization;
 674         bcopy(name, record->data->entry.name, bytes);
 675
 676         ++ip->ino_data.nlinks;
 677         ip->ino_data.ctime = trans->time;
 678         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 679
 680         /*
 681          * Find an unused namekey.  Both the in-memory record tree and
 682          * the B-Tree are checked.  We do not want historically deleted
 683          * names to create a collision as our iteration space may be limited,
 684          * and since create_tid wouldn't match anyway an ASOF search
 685          * must be used to locate collisions.
 686          *
 687          * delete-visibility is set so pending deletions do not give us
 688          * a false-negative on our ability to use an iterator.
 689          *
 690          * The iterator must not rollover the key.  Directory keys only
 691          * use the positive key space.
 692          */
 693         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 694         cursor.key_beg = record->leaf.base;
 695         cursor.flags |= HAMMER_CURSOR_ASOF;
 696         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 697         cursor.asof = ip->obj_asof;
 698
 699         while (hammer_ip_lookup(&cursor) == 0) {
 700                 ++record->leaf.base.key;
 701                 KKASSERT(record->leaf.base.key > 0);
 702                 cursor.key_beg.key = record->leaf.base.key;
 703                 if (--max_iterations == 0) {
 704                         hammer_rel_mem_record(record);
 705                         error = ENOSPC;
 706                         goto failed;
 707                 }
 708         }
 709
 710         /*
 711          * The target inode and the directory entry are bound together.
 712          */
 713         record->target_ip = ip;
 714         record->flush_state = HAMMER_FST_SETUP;
 715         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 716
 717         /*
 718          * The inode now has a dependancy and must be taken out of the idle
 719          * state.  An inode not in an idle state is given an extra reference.
 720          *
 721          * When transitioning to a SETUP state flag for an automatic reflush
 722          * when the dependancies are disposed of if someone is waiting on
 723          * the inode.
 724          */
 725         if (ip->flush_state == HAMMER_FST_IDLE) {
 726                 hammer_ref(&ip->lock);
 727                 ip->flush_state = HAMMER_FST_SETUP;
 728                 if (ip->flags & HAMMER_INODE_FLUSHW)
 729                         ip->flags |= HAMMER_INODE_REFLUSH;
 730         }
 731         error = hammer_mem_add(record);
 732         if (error == 0) {
 733                 dip->ino_data.mtime = trans->time;
 734                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 735         }
 736 failed:
 737         hammer_done_cursor(&cursor);
 738         return(error);
 739 }
 740
 741 /*
 742  * Delete the directory entry and update the inode link count.  The
 743  * cursor must be seeked to the directory entry record being deleted.
 744  *
 745  * The related inode should be share-locked by the caller.  The caller is
 746  * on the frontend.  It could also be NULL indicating that the directory
 747  * entry being removed has no related inode.
 748  *
 749  * This function can return EDEADLK requiring the caller to terminate
 750  * the cursor, any locks, wait on the returned record, and retry.
 751  */
 752 int
 753 hammer_ip_del_directory(struct hammer_transaction *trans,
 754                      hammer_cursor_t cursor, struct hammer_inode *dip,
 755                      struct hammer_inode *ip)
 756 {
 757         hammer_record_t record;
 758         int error;
 759
 760         if (hammer_cursor_inmem(cursor)) {
 761                 /*
 762                  * In-memory (unsynchronized) records can simply be freed.
 763                  *
 764                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 765                  * by the backend, we must still avoid races against the
 766                  * backend potentially syncing the record to the media.
 767                  *
 768                  * We cannot call hammer_ip_delete_record(), that routine may
 769                  * only be called from the backend.
 770                  */
 771                 record = cursor->iprec;
 772                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 773                                      HAMMER_RECF_DELETED_BE |
 774                                      HAMMER_RECF_COMMITTED)) {
 775                         KKASSERT(cursor->deadlk_rec == NULL);
 776                         hammer_ref(&record->lock);
 777                         cursor->deadlk_rec = record;
 778                         error = EDEADLK;
 779                 } else {
 780                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 781                         record->flags |= HAMMER_RECF_DELETED_FE;
 782                         error = 0;
 783                 }
 784         } else {
 785                 /*
 786                  * If the record is on-disk we have to queue the deletion by
 787                  * the record's key.  This also causes lookups to skip the
 788                  * record.
 789                  */
 790                 KKASSERT(dip->flags &
 791                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 792                 record = hammer_alloc_mem_record(dip, 0);
 793                 record->type = HAMMER_MEM_RECORD_DEL;
 794                 record->leaf.base = cursor->leaf->base;
 795
 796                 /*
 797                  * ip may be NULL, indicating the deletion of a directory
 798                  * entry which has no related inode.
 799                  */
 800                 record->target_ip = ip;
 801                 if (ip) {
 802                         record->flush_state = HAMMER_FST_SETUP;
 803                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 804                                           target_entry);
 805                 } else {
 806                         record->flush_state = HAMMER_FST_IDLE;
 807                 }
 808
 809                 /*
 810                  * The inode now has a dependancy and must be taken out of
 811                  * the idle state.  An inode not in an idle state is given
 812                  * an extra reference.
 813                  *
 814                  * When transitioning to a SETUP state flag for an automatic
 815                  * reflush when the dependancies are disposed of if someone
 816                  * is waiting on the inode.
 817                  */
 818                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 819                         hammer_ref(&ip->lock);
 820                         ip->flush_state = HAMMER_FST_SETUP;
 821                         if (ip->flags & HAMMER_INODE_FLUSHW)
 822                                 ip->flags |= HAMMER_INODE_REFLUSH;
 823                 }
 824
 825                 error = hammer_mem_add(record);
 826         }
 827
 828         /*
 829          * One less link.  The file may still be open in the OS even after
 830          * all links have gone away.
 831          *
 832          * We have to terminate the cursor before syncing the inode to
 833          * avoid deadlocking against ourselves.  XXX this may no longer
 834          * be true.
 835          *
 836          * If nlinks drops to zero and the vnode is inactive (or there is
 837          * no vnode), call hammer_inode_unloadable_check() to zonk the
 838          * inode.  If we don't do this here the inode will not be destroyed
 839          * on-media until we unmount.
 840          */
 841         if (error == 0) {
 842                 if (ip) {
 843                         --ip->ino_data.nlinks;  /* do before we might block */
 844                         ip->ino_data.ctime = trans->time;
 845                 }
 846                 dip->ino_data.mtime = trans->time;
 847                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 848                 if (ip) {
 849                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 850                         if (ip->ino_data.nlinks == 0 &&
 851                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 852                                 hammer_done_cursor(cursor);
 853                                 hammer_inode_unloadable_check(ip, 1);
 854                                 hammer_flush_inode(ip, 0);
 855                         }
 856                 }
 857
 858         }
 859         return(error);
 860 }
 861
 862 /*
 863  * Add a record to an inode.
 864  *
 865  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 866  * initialize the following additional fields:
 867  *
 868  * The related inode should be share-locked by the caller.  The caller is
 869  * on the frontend.
 870  *
 871  * record->rec.entry.base.base.key
 872  * record->rec.entry.base.base.rec_type
 873  * record->rec.entry.base.base.data_len
 874  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 875  */
 876 int
 877 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 878 {
 879         hammer_inode_t ip = record->ip;
 880         int error;
 881
 882         KKASSERT(record->leaf.base.localization != 0);
 883         record->leaf.base.obj_id = ip->obj_id;
 884         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 885         error = hammer_mem_add(record);
 886         return(error);
 887 }
 888
 889 /*
 890  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 891  * reserved so the front-end can flush large data writes without having
 892  * to queue the BIO to the flusher.  Only the related record gets queued
 893  * to the flusher.
 894  */
 895
 896 static hammer_record_t
 897 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 898 {
 899         struct hammer_bulk_info info;
 900
 901         bzero(&info, sizeof(info));
 902         info.leaf.base.obj_id = ip->obj_id;
 903         info.leaf.base.key = file_offset + bytes;
 904         info.leaf.base.create_tid = 0;
 905         info.leaf.base.delete_tid = 0;
 906         info.leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 907         info.leaf.base.obj_type = 0;                            /* unused */
 908         info.leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;        /* unused */
 909         info.leaf.base.localization = ip->obj_localization +    /* unused */
 910                                       HAMMER_LOCALIZE_MISC;
 911         info.leaf.data_len = bytes;
 912
 913         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 914                                    hammer_bulk_scan_callback, &info);
 915
 916         return(info.record);    /* may be NULL */
 917 }
 918
 919 /*
 920  * Take records vetted by overlap_cmp.  The first non-deleted record
 921  * (if any) stops the scan.
 922  */
 923 static int
 924 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 925 {
 926         struct hammer_bulk_info *info = data;
 927
 928         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 929                              HAMMER_RECF_COMMITTED)) {
 930                 return(0);
 931         }
 932         hammer_ref(&record->lock);
 933         info->record = record;
 934         return(-1);                     /* stop scan */
 935 }
 936
 937 /*
 938  * Reserve blockmap space placemarked with an in-memory record.
 939  *
 940  * This routine is called by the frontend in order to be able to directly
 941  * flush a buffer cache buffer.  The frontend has locked the related buffer
 942  * cache buffers and we should be able to manipulate any overlapping
 943  * in-memory records.
 944  *
 945  * The caller is responsible for adding the returned record.
 946  */
 947 hammer_record_t
 948 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 949                    int *errorp)
 950 {
 951         hammer_record_t record;
 952         hammer_record_t conflict;
 953         int zone;
 954
 955         /*
 956          * Deal with conflicting in-memory records.  We cannot have multiple
 957          * in-memory records for the same base offset without seriously
 958          * confusing the backend, including but not limited to the backend
 959          * issuing delete-create-delete or create-delete-create sequences
 960          * and asserting on the delete_tid being the same as the create_tid.
 961          *
 962          * If we encounter a record with the backend interlock set we cannot
 963          * immediately delete it without confusing the backend.
 964          */
 965         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 966                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 967                         conflict->flags |= HAMMER_RECF_WANTED;
 968                         tsleep(conflict, 0, "hmrrc3", 0);
 969                 } else {
 970                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 971                 }
 972                 hammer_rel_mem_record(conflict);
 973         }
 974
 975         /*
 976          * Create a record to cover the direct write.  This is called with
 977          * the related BIO locked so there should be no possible conflict.
 978          *
 979          * The backend is responsible for finalizing the space reserved in
 980          * this record.
 981          *
 982          * XXX bytes not aligned, depend on the reservation code to
 983          * align the reservation.
 984          */
 985         record = hammer_alloc_mem_record(ip, 0);
 986         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 987                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 988         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 989                                                &record->leaf.data_offset,
 990                                                errorp);
 991         if (record->resv == NULL) {
 992                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 993                 hammer_rel_mem_record(record);
 994                 return(NULL);
 995         }
 996         record->type = HAMMER_MEM_RECORD_DATA;
 997         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 998         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 999         record->leaf.base.obj_id = ip->obj_id;
1000         record->leaf.base.key = file_offset + bytes;
1001         record->leaf.base.localization = ip->obj_localization +
1002                                          HAMMER_LOCALIZE_MISC;
1003         record->leaf.data_len = bytes;
1004         hammer_crc_set_leaf(data, &record->leaf);
1005         KKASSERT(*errorp == 0);
1006         return(record);
1007 }
1008
1009 /*
1010  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1011  * and records in a flushing state are handled by the backend.  The vnops
1012  * setattr code will handle the block containing the truncation point.
1013  *
1014  * Partial blocks are not deleted.
1015  */
1016 int
1017 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1018 {
1019         struct rec_trunc_info info;
1020
1021         switch(ip->ino_data.obj_type) {
1022         case HAMMER_OBJTYPE_REGFILE:
1023                 info.rec_type = HAMMER_RECTYPE_DATA;
1024                 break;
1025         case HAMMER_OBJTYPE_DBFILE:
1026                 info.rec_type = HAMMER_RECTYPE_DB;
1027                 break;
1028         default:
1029                 return(EINVAL);
1030         }
1031         info.trunc_off = file_size;
1032         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1033                                    hammer_frontend_trunc_callback, &info);
1034         return(0);
1035 }
1036
1037 static int
1038 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1039 {
1040         if (record->flags & HAMMER_RECF_DELETED_FE)
1041                 return(0);
1042         if (record->flush_state == HAMMER_FST_FLUSH)
1043                 return(0);
1044         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1045         hammer_ref(&record->lock);
1046         record->flags |= HAMMER_RECF_DELETED_FE;
1047         hammer_rel_mem_record(record);
1048         return(0);
1049 }
1050
1051 /*
1052  * Return 1 if the caller must check for and delete existing records
1053  * before writing out a new data record.
1054  *
1055  * Return 0 if the caller can just insert the record into the B-Tree without
1056  * checking.
1057  */
1058 static int
1059 hammer_record_needs_overwrite_delete(hammer_record_t record)
1060 {
1061         hammer_inode_t ip = record->ip;
1062         int64_t file_offset;
1063         int r;
1064
1065         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1066                 file_offset = record->leaf.base.key;
1067         else
1068                 file_offset = record->leaf.base.key - record->leaf.data_len;
1069         r = (file_offset < ip->save_trunc_off);
1070         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1071                 if (ip->save_trunc_off <= record->leaf.base.key)
1072                         ip->save_trunc_off = record->leaf.base.key + 1;
1073         } else {
1074                 if (ip->save_trunc_off < record->leaf.base.key)
1075                         ip->save_trunc_off = record->leaf.base.key;
1076         }
1077         return(r);
1078 }
1079
1080 /*
1081  * Backend code.  Sync a record to the media.
1082  */
1083 int
1084 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1085 {
1086         hammer_transaction_t trans = cursor->trans;
1087         int64_t file_offset;
1088         int bytes;
1089         void *bdata;
1090         int error;
1091         int doprop;
1092
1093         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1094         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1095         KKASSERT(record->leaf.base.localization != 0);
1096
1097         /*
1098          * Any direct-write related to the record must complete before we
1099          * can sync the record to the on-disk media.
1100          */
1101         if (record->flags & (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL))
1102                 hammer_io_direct_wait(record);
1103
1104         /*
1105          * If this is a bulk-data record placemarker there may be an existing
1106          * record on-disk, indicating a data overwrite.  If there is the
1107          * on-disk record must be deleted before we can insert our new record.
1108          *
1109          * We've synthesized this record and do not know what the create_tid
1110          * on-disk is, nor how much data it represents.
1111          *
1112          * Keep in mind that (key) for data records is (base_offset + len),
1113          * not (base_offset).  Also, we only want to get rid of on-disk
1114          * records since we are trying to sync our in-memory record, call
1115          * hammer_ip_delete_range() with truncating set to 1 to make sure
1116          * it skips in-memory records.
1117          *
1118          * It is ok for the lookup to return ENOENT.
1119          *
1120          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1121          * to call hammer_ip_delete_range() or not.  This also means we must
1122          * update sync_trunc_off() as we write.
1123          */
1124         if (record->type == HAMMER_MEM_RECORD_DATA &&
1125             hammer_record_needs_overwrite_delete(record)) {
1126                 file_offset = record->leaf.base.key - record->leaf.data_len;
1127                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1128                         ~HAMMER_BUFMASK;
1129                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1130                 error = hammer_ip_delete_range(
1131                                 cursor, record->ip,
1132                                 file_offset, file_offset + bytes - 1,
1133                                 1);
1134                 if (error && error != ENOENT)
1135                         goto done;
1136         }
1137
1138         /*
1139          * If this is a general record there may be an on-disk version
1140          * that must be deleted before we can insert the new record.
1141          */
1142         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1143                 error = hammer_delete_general(cursor, record->ip,
1144                                               &record->leaf);
1145                 if (error && error != ENOENT)
1146                         goto done;
1147         }
1148
1149         /*
1150          * Setup the cursor.
1151          */
1152         hammer_normalize_cursor(cursor);
1153         cursor->key_beg = record->leaf.base;
1154         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1155         cursor->flags |= HAMMER_CURSOR_BACKEND;
1156         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1157
1158         /*
1159          * Records can wind up on-media before the inode itself is on-media.
1160          * Flag the case.
1161          */
1162         record->ip->flags |= HAMMER_INODE_DONDISK;
1163
1164         /*
1165          * If we are deleting a directory entry an exact match must be
1166          * found on-disk.
1167          */
1168         if (record->type == HAMMER_MEM_RECORD_DEL) {
1169                 error = hammer_btree_lookup(cursor);
1170                 if (error == 0) {
1171                         KKASSERT(cursor->iprec == NULL);
1172                         error = hammer_ip_delete_record(cursor, record->ip,
1173                                                         trans->tid);
1174                         if (error == 0) {
1175                                 record->flags |= HAMMER_RECF_DELETED_BE |
1176                                                  HAMMER_RECF_COMMITTED;
1177                                 ++record->ip->rec_generation;
1178                         }
1179                 }
1180                 goto done;
1181         }
1182
1183         /*
1184          * We are inserting.
1185          *
1186          * Issue a lookup to position the cursor and locate the insertion
1187          * point.  The target key should not exist.  If we are creating a
1188          * directory entry we may have to iterate the low 32 bits of the
1189          * key to find an unused key.
1190          */
1191         hammer_sync_lock_sh(trans);
1192         cursor->flags |= HAMMER_CURSOR_INSERT;
1193         error = hammer_btree_lookup(cursor);
1194         if (hammer_debug_inode)
1195                 kprintf("DOINSERT LOOKUP %d\n", error);
1196         if (error == 0) {
1197                 kprintf("hammer_ip_sync_record: duplicate rec "
1198                         "at (%016llx)\n", (long long)record->leaf.base.key);
1199                 Debugger("duplicate record1");
1200                 error = EIO;
1201         }
1202 #if 0
1203         if (record->type == HAMMER_MEM_RECORD_DATA)
1204                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1205                         record->leaf.base.key - record->leaf.data_len,
1206                         record->leaf.data_offset, error);
1207 #endif
1208
1209         if (error != ENOENT)
1210                 goto done_unlock;
1211
1212         /*
1213          * Allocate the record and data.  The result buffers will be
1214          * marked as being modified and further calls to
1215          * hammer_modify_buffer() will result in unneeded UNDO records.
1216          *
1217          * Support zero-fill records (data == NULL and data_len != 0)
1218          */
1219         if (record->type == HAMMER_MEM_RECORD_DATA) {
1220                 /*
1221                  * The data portion of a bulk-data record has already been
1222                  * committed to disk, we need only adjust the layer2
1223                  * statistics in the same transaction as our B-Tree insert.
1224                  */
1225                 KKASSERT(record->leaf.data_offset != 0);
1226                 error = hammer_blockmap_finalize(trans,
1227                                                  record->resv,
1228                                                  record->leaf.data_offset,
1229                                                  record->leaf.data_len);
1230         } else if (record->data && record->leaf.data_len) {
1231                 /*
1232                  * Wholely cached record, with data.  Allocate the data.
1233                  */
1234                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1235                                           record->leaf.base.rec_type,
1236                                           &record->leaf.data_offset,
1237                                           &cursor->data_buffer,
1238                                           0, &error);
1239                 if (bdata == NULL)
1240                         goto done_unlock;
1241                 hammer_crc_set_leaf(record->data, &record->leaf);
1242                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1243                 bcopy(record->data, bdata, record->leaf.data_len);
1244                 hammer_modify_buffer_done(cursor->data_buffer);
1245         } else {
1246                 /*
1247                  * Wholely cached record, without data.
1248                  */
1249                 record->leaf.data_offset = 0;
1250                 record->leaf.data_crc = 0;
1251         }
1252
1253         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1254         if (hammer_debug_inode && error) {
1255                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1256                         error,
1257                         (long long)cursor->node->node_offset,
1258                         cursor->index,
1259                         (long long)record->leaf.base.key);
1260         }
1261
1262         /*
1263          * Our record is on-disk and we normally mark the in-memory version
1264          * as having been committed (and not BE-deleted).
1265          *
1266          * If the record represented a directory deletion but we had to
1267          * sync a valid directory entry to disk due to dependancies,
1268          * we must convert the record to a covering delete so the
1269          * frontend does not have visibility on the synced entry.
1270          */
1271         if (error == 0) {
1272                 if (doprop) {
1273                         hammer_btree_do_propagation(cursor,
1274                                                     record->ip->pfsm,
1275                                                     &record->leaf);
1276                 }
1277                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1278                         /*
1279                          * Must convert deleted directory entry add
1280                          * to a directory entry delete.
1281                          */
1282                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1283                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1284                         record->type = HAMMER_MEM_RECORD_DEL;
1285                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1286                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1287                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1288                                                  HAMMER_RECF_DELETED_BE)) == 0);
1289                         /* converted record is not yet committed */
1290                         /* hammer_flush_record_done takes care of the rest */
1291                 } else {
1292                         /*
1293                          * Everything went fine and we are now done with
1294                          * this record.
1295                          */
1296                         record->flags |= HAMMER_RECF_COMMITTED;
1297                         ++record->ip->rec_generation;
1298                 }
1299         } else {
1300                 if (record->leaf.data_offset) {
1301                         hammer_blockmap_free(trans, record->leaf.data_offset,
1302                                              record->leaf.data_len);
1303                 }
1304         }
1305 done_unlock:
1306         hammer_sync_unlock(trans);
1307 done:
1308         return(error);
1309 }
1310
1311 /*
1312  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1313  * entry's key is used to deal with hash collisions in the upper 32 bits.
1314  * A unique 64 bit key is generated in-memory and may be regenerated a
1315  * second time when the directory record is flushed to the on-disk B-Tree.
1316  *
1317  * A referenced record is passed to this function.  This function
1318  * eats the reference.  If an error occurs the record will be deleted.
1319  *
1320  * A copy of the temporary record->data pointer provided by the caller
1321  * will be made.
1322  */
1323 int
1324 hammer_mem_add(hammer_record_t record)
1325 {
1326         hammer_mount_t hmp = record->ip->hmp;
1327
1328         /*
1329          * Make a private copy of record->data
1330          */
1331         if (record->data)
1332                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1333
1334         /*
1335          * Insert into the RB tree.  A unique key should have already
1336          * been selected if this is a directory entry.
1337          */
1338         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1339                 record->flags |= HAMMER_RECF_DELETED_FE;
1340                 hammer_rel_mem_record(record);
1341                 return (EEXIST);
1342         }
1343         ++hmp->count_newrecords;
1344         ++hmp->rsv_recs;
1345         ++record->ip->rsv_recs;
1346         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1347         record->flags |= HAMMER_RECF_ONRBTREE;
1348         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1349         hammer_rel_mem_record(record);
1350         return(0);
1351 }
1352
1353 /************************************************************************
1354  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1355  ************************************************************************
1356  *
1357  * These functions augment the B-Tree scanning functions in hammer_btree.c
1358  * by merging in-memory records with on-disk records.
1359  */
1360
1361 /*
1362  * Locate a particular record either in-memory or on-disk.
1363  *
1364  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1365  * NOT be called to iterate results.
1366  */
1367 int
1368 hammer_ip_lookup(hammer_cursor_t cursor)
1369 {
1370         int error;
1371
1372         /*
1373          * If the element is in-memory return it without searching the
1374          * on-disk B-Tree
1375          */
1376         KKASSERT(cursor->ip);
1377         error = hammer_mem_lookup(cursor);
1378         if (error == 0) {
1379                 cursor->leaf = &cursor->iprec->leaf;
1380                 return(error);
1381         }
1382         if (error != ENOENT)
1383                 return(error);
1384
1385         /*
1386          * If the inode has on-disk components search the on-disk B-Tree.
1387          */
1388         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1389                 return(error);
1390         error = hammer_btree_lookup(cursor);
1391         if (error == 0)
1392                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1393         return(error);
1394 }
1395
1396 /*
1397  * Helper for hammer_ip_first()/hammer_ip_next()
1398  *
1399  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1400  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1401  * state if hammer_ip_next() needs to re-seek.
1402  */
1403 static __inline
1404 int
1405 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1406 {
1407         hammer_inode_t ip = cursor->ip;
1408         int error;
1409
1410         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1411                 error = hammer_btree_lookup(cursor);
1412                 if (error == ENOENT || error == EDEADLK) {
1413                         if (hammer_debug_general & 0x2000) {
1414                                 kprintf("error %d node %p %016llx index %d\n",
1415                                         error, cursor->node,
1416                                         (long long)cursor->node->node_offset,
1417                                         cursor->index);
1418                         }
1419                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1420                         error = hammer_btree_iterate(cursor);
1421                 }
1422                 if (error == 0) {
1423                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1424                                            HAMMER_CURSOR_ATEDISK);
1425                 } else {
1426                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1427                                          HAMMER_CURSOR_ATEDISK;
1428                         if (error == ENOENT)
1429                                 error = 0;
1430                 }
1431         } else {
1432                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1433                 error = 0;
1434         }
1435         return(error);
1436 }
1437
1438 /*
1439  * Helper for hammer_ip_next()
1440  *
1441  * The caller has determined that the media cursor is further along than the
1442  * memory cursor and must be reseeked after a generation number change.
1443  */
1444 static
1445 int
1446 _hammer_ip_reseek(hammer_cursor_t cursor)
1447 {
1448         struct hammer_base_elm save;
1449         hammer_btree_elm_t elm;
1450         int error;
1451         int r;
1452         int again = 0;
1453
1454         /*
1455          * Do the re-seek.
1456          */
1457         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1458                 (long long)cursor->ip->obj_id);
1459         save = cursor->key_beg;
1460         cursor->key_beg = cursor->iprec->leaf.base;
1461         error = _hammer_ip_seek_btree(cursor);
1462         KKASSERT(error == 0);
1463         cursor->key_beg = save;
1464
1465         /*
1466          * If the memory record was previous returned to
1467          * the caller and the media record matches
1468          * (-1/+1: only create_tid differs), then iterate
1469          * the media record to avoid a double result.
1470          */
1471         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1472             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1473                 elm = &cursor->node->ondisk->elms[cursor->index];
1474                 r = hammer_btree_cmp(&elm->base,
1475                                      &cursor->iprec->leaf.base);
1476                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1477                         if (r >= -1 && r <= 1) {
1478                                 kprintf("HAMMER: Debug: iterated after "
1479                                         "re-seek (asof r=%d)\n", r);
1480                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1481                                 again = 1;
1482                         }
1483                 } else {
1484                         if (r == 0) {
1485                                 kprintf("HAMMER: Debug: iterated after "
1486                                         "re-seek\n");
1487                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1488                                 again = 1;
1489                         }
1490                 }
1491         }
1492         return(again);
1493 }
1494
1495 /*
1496  * Locate the first record within the cursor's key_beg/key_end range,
1497  * restricted to a particular inode.  0 is returned on success, ENOENT
1498  * if no records matched the requested range, or some other error.
1499  *
1500  * When 0 is returned hammer_ip_next() may be used to iterate additional
1501  * records within the requested range.
1502  *
1503  * This function can return EDEADLK, requiring the caller to terminate
1504  * the cursor and try again.
1505  */
1506
1507 int
1508 hammer_ip_first(hammer_cursor_t cursor)
1509 {
1510         hammer_inode_t ip = cursor->ip;
1511         int error;
1512
1513         KKASSERT(ip != NULL);
1514
1515         /*
1516          * Clean up fields and setup for merged scan
1517          */
1518         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1519
1520         /*
1521          * Search the in-memory record list (Red-Black tree).  Unlike the
1522          * B-Tree search, mem_first checks for records in the range.
1523          *
1524          * This function will setup both ATEMEM and MEMEOF properly for
1525          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1526          */
1527         hammer_mem_first(cursor);
1528
1529         /*
1530          * Detect generation changes during blockages, including
1531          * blockages which occur on the initial btree search.
1532          */
1533         cursor->rec_generation = cursor->ip->rec_generation;
1534
1535         /*
1536          * Initial search and result
1537          */
1538         error = _hammer_ip_seek_btree(cursor);
1539         if (error == 0)
1540                 error = hammer_ip_next(cursor);
1541
1542         return (error);
1543 }
1544
1545 /*
1546  * Retrieve the next record in a merged iteration within the bounds of the
1547  * cursor.  This call may be made multiple times after the cursor has been
1548  * initially searched with hammer_ip_first().
1549  *
1550  * There are numerous special cases in this code to deal with races between
1551  * in-memory records and on-media records.
1552  *
1553  * 0 is returned on success, ENOENT if no further records match the
1554  * requested range, or some other error code is returned.
1555  */
1556 int
1557 hammer_ip_next(hammer_cursor_t cursor)
1558 {
1559         hammer_btree_elm_t elm;
1560         hammer_record_t rec;
1561         hammer_record_t tmprec;
1562         int error;
1563         int r;
1564
1565 again:
1566         /*
1567          * Get the next on-disk record
1568          *
1569          * NOTE: If we deleted the last on-disk record we had scanned
1570          *       ATEDISK will be clear and RETEST will be set, forcing
1571          *       a call to iterate.  The fact that ATEDISK is clear causes
1572          *       iterate to re-test the 'current' element.  If ATEDISK is
1573          *       set, iterate will skip the 'current' element.
1574          */
1575         error = 0;
1576         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1577                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1578                                      HAMMER_CURSOR_RETEST)) {
1579                         error = hammer_btree_iterate(cursor);
1580                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1581                         if (error == 0) {
1582                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1583                                 hammer_cache_node(&cursor->ip->cache[1],
1584                                                   cursor->node);
1585                         } else if (error == ENOENT) {
1586                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1587                                                  HAMMER_CURSOR_ATEDISK;
1588                                 error = 0;
1589                         }
1590                 }
1591         }
1592
1593         /*
1594          * If the generation changed the backend has deleted or committed
1595          * one or more memory records since our last check.
1596          *
1597          * When this case occurs if the disk cursor is > current memory record
1598          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1599          * Since the cursor is ahead it must have not yet been eaten (if
1600          * not at eof anyway). (XXX data offset case?)
1601          *
1602          * NOTE: we are not doing a full check here.  That will be handled
1603          * later on.
1604          *
1605          * If we have exhausted all memory records we do not have to do any
1606          * further seeks.
1607          */
1608         while (cursor->rec_generation != cursor->ip->rec_generation &&
1609                error == 0
1610         ) {
1611                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1612                 cursor->rec_generation = cursor->ip->rec_generation;
1613                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1614                         break;
1615                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1616                         r = 1;
1617                 } else {
1618                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1619                         elm = &cursor->node->ondisk->elms[cursor->index];
1620                         r = hammer_btree_cmp(&elm->base,
1621                                              &cursor->iprec->leaf.base);
1622                 }
1623
1624                 /*
1625                  * Do we re-seek the media cursor?
1626                  */
1627                 if (r > 0) {
1628                         if (_hammer_ip_reseek(cursor))
1629                                 goto again;
1630                 }
1631         }
1632
1633         /*
1634          * We can now safely get the next in-memory record.  We cannot
1635          * block here.
1636          *
1637          * hammer_rec_scan_cmp:  Is the record still in our general range,
1638          *                       (non-inclusive of snapshot exclusions)?
1639          * hammer_rec_scan_callback: Is the record in our snapshot?
1640          */
1641         tmprec = NULL;
1642         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1643                 /*
1644                  * If the current memory record was eaten then get the next
1645                  * one.  Stale records are skipped.
1646                  */
1647                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1648                         tmprec = cursor->iprec;
1649                         cursor->iprec = NULL;
1650                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1651                         while (rec) {
1652                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1653                                         break;
1654                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1655                                         break;
1656                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1657                         }
1658                         if (cursor->iprec) {
1659                                 KKASSERT(cursor->iprec == rec);
1660                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1661                         } else {
1662                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1663                         }
1664                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1665                 }
1666         }
1667
1668         /*
1669          * MEMORY RECORD VALIDITY TEST
1670          *
1671          * (We still can't block, which is why tmprec is being held so
1672          * long).
1673          *
1674          * If the memory record is no longer valid we skip it.  It may
1675          * have been deleted by the frontend.  If it was deleted or
1676          * committed by the backend the generation change re-seeked the
1677          * disk cursor and the record will be present there.
1678          */
1679         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1680                 KKASSERT(cursor->iprec);
1681                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1682                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1683                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1684                         if (tmprec)
1685                                 hammer_rel_mem_record(tmprec);
1686                         goto again;
1687                 }
1688         }
1689         if (tmprec)
1690                 hammer_rel_mem_record(tmprec);
1691
1692         /*
1693          * Extract either the disk or memory record depending on their
1694          * relative position.
1695          */
1696         error = 0;
1697         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1698         case 0:
1699                 /*
1700                  * Both entries valid.   Compare the entries and nominally
1701                  * return the first one in the sort order.  Numerous cases
1702                  * require special attention, however.
1703                  */
1704                 elm = &cursor->node->ondisk->elms[cursor->index];
1705                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1706
1707                 /*
1708                  * If the two entries differ only by their key (-2/2) or
1709                  * create_tid (-1/1), and are DATA records, we may have a
1710                  * nominal match.  We have to calculate the base file
1711                  * offset of the data.
1712                  */
1713                 if (r <= 2 && r >= -2 && r != 0 &&
1714                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1715                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1716                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1717                         int64_t base2 = cursor->iprec->leaf.base.key -
1718                                         cursor->iprec->leaf.data_len;
1719                         if (base1 == base2)
1720                                 r = 0;
1721                 }
1722
1723                 if (r < 0) {
1724                         error = hammer_btree_extract(cursor,
1725                                                      HAMMER_CURSOR_GET_LEAF);
1726                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1727                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1728                         break;
1729                 }
1730
1731                 /*
1732                  * If the entries match exactly the memory entry is either
1733                  * an on-disk directory entry deletion or a bulk data
1734                  * overwrite.  If it is a directory entry deletion we eat
1735                  * both entries.
1736                  *
1737                  * For the bulk-data overwrite case it is possible to have
1738                  * visibility into both, which simply means the syncer
1739                  * hasn't gotten around to doing the delete+insert sequence
1740                  * on the B-Tree.  Use the memory entry and throw away the
1741                  * on-disk entry.
1742                  *
1743                  * If the in-memory record is not either of these we
1744                  * probably caught the syncer while it was syncing it to
1745                  * the media.  Since we hold a shared lock on the cursor,
1746                  * the in-memory record had better be marked deleted at
1747                  * this point.
1748                  */
1749                 if (r == 0) {
1750                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1751                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1752                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1753                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1754                                         goto again;
1755                                 }
1756                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1757                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1758                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1759                                 }
1760                                 /* fall through to memory entry */
1761                         } else {
1762                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1763                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1764                                 goto again;
1765                         }
1766                 }
1767                 /* fall through to the memory entry */
1768         case HAMMER_CURSOR_ATEDISK:
1769                 /*
1770                  * Only the memory entry is valid.
1771                  */
1772                 cursor->leaf = &cursor->iprec->leaf;
1773                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1774                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1775
1776                 /*
1777                  * If the memory entry is an on-disk deletion we should have
1778                  * also had found a B-Tree record.  If the backend beat us
1779                  * to it it would have interlocked the cursor and we should
1780                  * have seen the in-memory record marked DELETED_FE.
1781                  */
1782                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1783                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1784                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1785                 }
1786                 break;
1787         case HAMMER_CURSOR_ATEMEM:
1788                 /*
1789                  * Only the disk entry is valid
1790                  */
1791                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1792                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1793                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1794                 break;
1795         default:
1796                 /*
1797                  * Neither entry is valid
1798                  *
1799                  * XXX error not set properly
1800                  */
1801                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1802                 cursor->leaf = NULL;
1803                 error = ENOENT;
1804                 break;
1805         }
1806         return(error);
1807 }
1808
1809 /*
1810  * Resolve the cursor->data pointer for the current cursor position in
1811  * a merged iteration.
1812  */
1813 int
1814 hammer_ip_resolve_data(hammer_cursor_t cursor)
1815 {
1816         hammer_record_t record;
1817         int error;
1818
1819         if (hammer_cursor_inmem(cursor)) {
1820                 /*
1821                  * The data associated with an in-memory record is usually
1822                  * kmalloced, but reserve-ahead data records will have an
1823                  * on-disk reference.
1824                  *
1825                  * NOTE: Reserve-ahead data records must be handled in the
1826                  * context of the related high level buffer cache buffer
1827                  * to interlock against async writes.
1828                  */
1829                 record = cursor->iprec;
1830                 cursor->data = record->data;
1831                 error = 0;
1832                 if (cursor->data == NULL) {
1833                         KKASSERT(record->leaf.base.rec_type ==
1834                                  HAMMER_RECTYPE_DATA);
1835                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1836                                                     record->leaf.data_offset,
1837                                                     record->leaf.data_len,
1838                                                     &error,
1839                                                     &cursor->data_buffer);
1840                 }
1841         } else {
1842                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1843                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1844         }
1845         return(error);
1846 }
1847
1848 /*
1849  * Backend truncation / record replacement - delete records in range.
1850  *
1851  * Delete all records within the specified range for inode ip.  In-memory
1852  * records still associated with the frontend are ignored.
1853  *
1854  * If truncating is non-zero in-memory records associated with the back-end
1855  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1856  *
1857  * NOTES:
1858  *
1859  *      * An unaligned range will cause new records to be added to cover
1860  *        the edge cases. (XXX not implemented yet).
1861  *
1862  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1863  *        also do not deal with unaligned ranges.
1864  *
1865  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1866  *
1867  *      * Record keys for regular file data have to be special-cased since
1868  *        they indicate the end of the range (key = base + bytes).
1869  *
1870  *      * This function may be asked to delete ridiculously huge ranges, for
1871  *        example if someone truncates or removes a 1TB regular file.  We
1872  *        must be very careful on restarts and we may have to stop w/
1873  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1874  */
1875 int
1876 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1877                        int64_t ran_beg, int64_t ran_end, int truncating)
1878 {
1879         hammer_transaction_t trans = cursor->trans;
1880         hammer_btree_leaf_elm_t leaf;
1881         int error;
1882         int64_t off;
1883         int64_t tmp64;
1884
1885 #if 0
1886         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1887 #endif
1888
1889         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1890 retry:
1891         hammer_normalize_cursor(cursor);
1892         cursor->key_beg.localization = ip->obj_localization +
1893                                        HAMMER_LOCALIZE_MISC;
1894         cursor->key_beg.obj_id = ip->obj_id;
1895         cursor->key_beg.create_tid = 0;
1896         cursor->key_beg.delete_tid = 0;
1897         cursor->key_beg.obj_type = 0;
1898
1899         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1900                 cursor->key_beg.key = ran_beg;
1901                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1902         } else {
1903                 /*
1904                  * The key in the B-Tree is (base+bytes), so the first possible
1905                  * matching key is ran_beg + 1.
1906                  */
1907                 cursor->key_beg.key = ran_beg + 1;
1908                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1909         }
1910
1911         cursor->key_end = cursor->key_beg;
1912         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1913                 cursor->key_end.key = ran_end;
1914         } else {
1915                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1916                 if (tmp64 < ran_end)
1917                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1918                 else
1919                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1920         }
1921
1922         cursor->asof = ip->obj_asof;
1923         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1924         cursor->flags |= HAMMER_CURSOR_ASOF;
1925         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1926         cursor->flags |= HAMMER_CURSOR_BACKEND;
1927         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1928
1929         error = hammer_ip_first(cursor);
1930
1931         /*
1932          * Iterate through matching records and mark them as deleted.
1933          */
1934         while (error == 0) {
1935                 leaf = cursor->leaf;
1936
1937                 KKASSERT(leaf->base.delete_tid == 0);
1938                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1939
1940                 /*
1941                  * There may be overlap cases for regular file data.  Also
1942                  * remember the key for a regular file record is (base + len),
1943                  * NOT (base).
1944                  *
1945                  * Note that do to duplicates (mem & media) allowed by
1946                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1947                  */
1948                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1949                         off = leaf->base.key - leaf->data_len;
1950                         /*
1951                          * Check the left edge case.  We currently do not
1952                          * split existing records.
1953                          */
1954                         if (off < ran_beg && leaf->base.key > ran_beg) {
1955                                 panic("hammer left edge case %016llx %d\n",
1956                                         (long long)leaf->base.key,
1957                                         leaf->data_len);
1958                         }
1959
1960                         /*
1961                          * Check the right edge case.  Note that the
1962                          * record can be completely out of bounds, which
1963                          * terminates the search.
1964                          *
1965                          * base->key is exclusive of the right edge while
1966                          * ran_end is inclusive of the right edge.  The
1967                          * (key - data_len) left boundary is inclusive.
1968                          *
1969                          * XXX theory-check this test at some point, are
1970                          * we missing a + 1 somewhere?  Note that ran_end
1971                          * could overflow.
1972                          */
1973                         if (leaf->base.key - 1 > ran_end) {
1974                                 if (leaf->base.key - leaf->data_len > ran_end)
1975                                         break;
1976                                 panic("hammer right edge case\n");
1977                         }
1978                 } else {
1979                         off = leaf->base.key;
1980                 }
1981
1982                 /*
1983                  * Delete the record.  When truncating we do not delete
1984                  * in-memory (data) records because they represent data
1985                  * written after the truncation.
1986                  *
1987                  * This will also physically destroy the B-Tree entry and
1988                  * data if the retention policy dictates.  The function
1989                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
1990                  * to retest the new 'current' element.
1991                  */
1992                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
1993                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
1994                         /*
1995                          * If we have built up too many meta-buffers we risk
1996                          * deadlocking the kernel and must stop.  This can
1997                          * occur when deleting ridiculously huge files.
1998                          * sync_trunc_off is updated so the next cycle does
1999                          * not re-iterate records we have already deleted.
2000                          *
2001                          * This is only done with formal truncations.
2002                          */
2003                         if (truncating > 1 && error == 0 &&
2004                             hammer_flusher_meta_limit(ip->hmp)) {
2005                                 ip->sync_trunc_off = off;
2006                                 error = EWOULDBLOCK;
2007                         }
2008                 }
2009                 if (error)
2010                         break;
2011                 ran_beg = off;  /* for restart */
2012                 error = hammer_ip_next(cursor);
2013         }
2014         if (cursor->node)
2015                 hammer_cache_node(&ip->cache[1], cursor->node);
2016
2017         if (error == EDEADLK) {
2018                 hammer_done_cursor(cursor);
2019                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2020                 if (error == 0)
2021                         goto retry;
2022         }
2023         if (error == ENOENT)
2024                 error = 0;
2025         return(error);
2026 }
2027
2028 /*
2029  * This backend function deletes the specified record on-disk, similar to
2030  * delete_range but for a specific record.  Unlike the exact deletions
2031  * used when deleting a directory entry this function uses an ASOF search
2032  * like delete_range.
2033  *
2034  * This function may be called with ip->obj_asof set for a slave snapshot,
2035  * so don't use it.  We always delete non-historical records only.
2036  */
2037 static int
2038 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2039                       hammer_btree_leaf_elm_t leaf)
2040 {
2041         hammer_transaction_t trans = cursor->trans;
2042         int error;
2043
2044         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2045 retry:
2046         hammer_normalize_cursor(cursor);
2047         cursor->key_beg = leaf->base;
2048         cursor->asof = HAMMER_MAX_TID;
2049         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2050         cursor->flags |= HAMMER_CURSOR_ASOF;
2051         cursor->flags |= HAMMER_CURSOR_BACKEND;
2052         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2053
2054         error = hammer_btree_lookup(cursor);
2055         if (error == 0) {
2056                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2057         }
2058         if (error == EDEADLK) {
2059                 hammer_done_cursor(cursor);
2060                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2061                 if (error == 0)
2062                         goto retry;
2063         }
2064         return(error);
2065 }
2066
2067 /*
2068  * This function deletes remaining auxillary records when an inode is
2069  * being deleted.  This function explicitly does not delete the
2070  * inode record, directory entry, data, or db records.  Those must be
2071  * properly disposed of prior to this call.
2072  */
2073 int
2074 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2075 {
2076         hammer_transaction_t trans = cursor->trans;
2077         hammer_btree_leaf_elm_t leaf;
2078         int error;
2079
2080         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2081 retry:
2082         hammer_normalize_cursor(cursor);
2083         cursor->key_beg.localization = ip->obj_localization +
2084                                        HAMMER_LOCALIZE_MISC;
2085         cursor->key_beg.obj_id = ip->obj_id;
2086         cursor->key_beg.create_tid = 0;
2087         cursor->key_beg.delete_tid = 0;
2088         cursor->key_beg.obj_type = 0;
2089         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2090         cursor->key_beg.key = HAMMER_MIN_KEY;
2091
2092         cursor->key_end = cursor->key_beg;
2093         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2094         cursor->key_end.key = HAMMER_MAX_KEY;
2095
2096         cursor->asof = ip->obj_asof;
2097         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2098         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2099         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2100         cursor->flags |= HAMMER_CURSOR_BACKEND;
2101
2102         error = hammer_ip_first(cursor);
2103
2104         /*
2105          * Iterate through matching records and mark them as deleted.
2106          */
2107         while (error == 0) {
2108                 leaf = cursor->leaf;
2109
2110                 KKASSERT(leaf->base.delete_tid == 0);
2111
2112                 /*
2113                  * Mark the record and B-Tree entry as deleted.  This will
2114                  * also physically delete the B-Tree entry, record, and
2115                  * data if the retention policy dictates.  The function
2116                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2117                  * to retest the new 'current' element.
2118                  *
2119                  * Directory entries (and delete-on-disk directory entries)
2120                  * must be synced and cannot be deleted.
2121                  */
2122                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2123                 ++*countp;
2124                 if (error)
2125                         break;
2126                 error = hammer_ip_next(cursor);
2127         }
2128         if (cursor->node)
2129                 hammer_cache_node(&ip->cache[1], cursor->node);
2130         if (error == EDEADLK) {
2131                 hammer_done_cursor(cursor);
2132                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2133                 if (error == 0)
2134                         goto retry;
2135         }
2136         if (error == ENOENT)
2137                 error = 0;
2138         return(error);
2139 }
2140
2141 /*
2142  * Delete the record at the current cursor.  On success the cursor will
2143  * be positioned appropriately for an iteration but may no longer be at
2144  * a leaf node.
2145  *
2146  * This routine is only called from the backend.
2147  *
2148  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2149  * cursor and retry.
2150  */
2151 int
2152 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2153                         hammer_tid_t tid)
2154 {
2155         hammer_record_t iprec;
2156         hammer_mount_t hmp;
2157         int error;
2158
2159         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2160         KKASSERT(tid != 0);
2161         hmp = cursor->node->hmp;
2162
2163         /*
2164          * In-memory (unsynchronized) records can simply be freed.  This
2165          * only occurs in range iterations since all other records are
2166          * individually synchronized.  Thus there should be no confusion with
2167          * the interlock.
2168          *
2169          * An in-memory record may be deleted before being committed to disk,
2170          * but could have been accessed in the mean time.  The reservation
2171          * code will deal with the case.
2172          */
2173         if (hammer_cursor_inmem(cursor)) {
2174                 iprec = cursor->iprec;
2175                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2176                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2177                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2178                 KKASSERT(iprec->ip == ip);
2179                 ++ip->rec_generation;
2180                 return(0);
2181         }
2182
2183         /*
2184          * On-disk records are marked as deleted by updating their delete_tid.
2185          * This does not effect their position in the B-Tree (which is based
2186          * on their create_tid).
2187          *
2188          * Frontend B-Tree operations track inodes so we tell
2189          * hammer_delete_at_cursor() not to.
2190          */
2191         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2192
2193         if (error == 0) {
2194                 error = hammer_delete_at_cursor(
2195                                 cursor,
2196                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2197                                 cursor->trans->tid,
2198                                 cursor->trans->time32,
2199                                 0, NULL);
2200         }
2201         return(error);
2202 }
2203
2204 /*
2205  * Delete the B-Tree element at the current cursor and do any necessary
2206  * mirror propagation.
2207  *
2208  * The cursor must be properly positioned for an iteration on return but
2209  * may be pointing at an internal element.
2210  *
2211  * An element can be un-deleted by passing a delete_tid of 0 with
2212  * HAMMER_DELETE_ADJUST.
2213  */
2214 int
2215 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2216                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2217                         int track, int64_t *stat_bytes)
2218 {
2219         struct hammer_btree_leaf_elm save_leaf;
2220         hammer_transaction_t trans;
2221         hammer_btree_leaf_elm_t leaf;
2222         hammer_node_t node;
2223         hammer_btree_elm_t elm;
2224         hammer_off_t data_offset;
2225         int32_t data_len;
2226         u_int16_t rec_type;
2227         int error;
2228         int icount;
2229         int doprop;
2230
2231         error = hammer_cursor_upgrade(cursor);
2232         if (error)
2233                 return(error);
2234
2235         trans = cursor->trans;
2236         node = cursor->node;
2237         elm = &node->ondisk->elms[cursor->index];
2238         leaf = &elm->leaf;
2239         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2240
2241         hammer_sync_lock_sh(trans);
2242         doprop = 0;
2243         icount = 0;
2244
2245         /*
2246          * Adjust the delete_tid.  Update the mirror_tid propagation field
2247          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2248          */
2249         if (delete_flags & HAMMER_DELETE_ADJUST) {
2250                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2251                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2252                                 icount = -1;
2253                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2254                                 icount = 1;
2255                 }
2256
2257                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2258                 elm->leaf.base.delete_tid = delete_tid;
2259                 elm->leaf.delete_ts = delete_ts;
2260                 hammer_modify_node_done(node);
2261
2262                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2263                         hammer_modify_node_field(trans, node, mirror_tid);
2264                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2265                         hammer_modify_node_done(node);
2266                         doprop = 1;
2267                         if (hammer_debug_general & 0x0002) {
2268                                 kprintf("delete_at_cursor: propagate %016llx"
2269                                         " @%016llx\n",
2270                                         (long long)elm->leaf.base.delete_tid,
2271                                         (long long)node->node_offset);
2272                         }
2273                 }
2274
2275                 /*
2276                  * Adjust for the iteration.  We have deleted the current
2277                  * element and want to clear ATEDISK so the iteration does
2278                  * not skip the element after, which now becomes the current
2279                  * element.  This element must be re-tested if doing an
2280                  * iteration, which is handled by the RETEST flag.
2281                  */
2282                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2283                         cursor->flags |= HAMMER_CURSOR_RETEST;
2284                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2285                 }
2286
2287                 /*
2288                  * An on-disk record cannot have the same delete_tid
2289                  * as its create_tid.  In a chain of record updates
2290                  * this could result in a duplicate record.
2291                  */
2292                 KKASSERT(elm->leaf.base.delete_tid !=
2293                          elm->leaf.base.create_tid);
2294         }
2295
2296         /*
2297          * Destroy the B-Tree element if asked (typically if a nohistory
2298          * file or mount, or when called by the pruning code).
2299          *
2300          * Adjust the ATEDISK flag to properly support iterations.
2301          */
2302         if (delete_flags & HAMMER_DELETE_DESTROY) {
2303                 data_offset = elm->leaf.data_offset;
2304                 data_len = elm->leaf.data_len;
2305                 rec_type = elm->leaf.base.rec_type;
2306                 if (doprop) {
2307                         save_leaf = elm->leaf;
2308                         leaf = &save_leaf;
2309                 }
2310                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2311                     elm->leaf.base.delete_tid == 0) {
2312                         icount = -1;
2313                 }
2314
2315                 error = hammer_btree_delete(cursor);
2316                 if (error == 0) {
2317                         /*
2318                          * The deletion moves the next element (if any) to
2319                          * the current element position.  We must clear
2320                          * ATEDISK so this element is not skipped and we
2321                          * must set RETEST to force any iteration to re-test
2322                          * the element.
2323                          */
2324                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2325                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2326                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2327                         }
2328                 }
2329                 if (error == 0) {
2330                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2331                         case HAMMER_ZONE_LARGE_DATA:
2332                         case HAMMER_ZONE_SMALL_DATA:
2333                         case HAMMER_ZONE_META:
2334                                 hammer_blockmap_free(trans,
2335                                                      data_offset, data_len);
2336                                 break;
2337                         default:
2338                                 break;
2339                         }
2340                 }
2341         }
2342
2343         /*
2344          * Track inode count and next_tid.  This is used by the mirroring
2345          * and PFS code.  icount can be negative, zero, or positive.
2346          */
2347         if (error == 0 && track) {
2348                 if (icount) {
2349                         hammer_modify_volume_field(trans, trans->rootvol,
2350                                                    vol0_stat_inodes);
2351                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2352                         hammer_modify_volume_done(trans->rootvol);
2353                 }
2354                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2355                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2356                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2357                         hammer_modify_volume_done(trans->rootvol);
2358                 }
2359         }
2360
2361         /*
2362          * mirror_tid propagation occurs if the node's mirror_tid had to be
2363          * updated while adjusting the delete_tid.
2364          *
2365          * This occurs when deleting even in nohistory mode, but does not
2366          * occur when pruning an already-deleted node.
2367          *
2368          * cursor->ip is NULL when called from the pruning, mirroring,
2369          * and pfs code.  If non-NULL propagation will be conditionalized
2370          * on whether the PFS is in no-history mode or not.
2371          */
2372         if (doprop) {
2373                 if (cursor->ip)
2374                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2375                 else
2376                         hammer_btree_do_propagation(cursor, NULL, leaf);
2377         }
2378         hammer_sync_unlock(trans);
2379         return (error);
2380 }
2381
2382 /*
2383  * Determine whether we can remove a directory.  This routine checks whether
2384  * a directory is empty or not and enforces flush connectivity.
2385  *
2386  * Flush connectivity requires that we block if the target directory is
2387  * currently flushing, otherwise it may not end up in the same flush group.
2388  *
2389  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2390  */
2391 int
2392 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2393 {
2394         struct hammer_cursor cursor;
2395         int error;
2396
2397         /*
2398          * Check directory empty
2399          */
2400         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2401
2402         cursor.key_beg.localization = ip->obj_localization +
2403                                       hammer_dir_localization(ip);
2404         cursor.key_beg.obj_id = ip->obj_id;
2405         cursor.key_beg.create_tid = 0;
2406         cursor.key_beg.delete_tid = 0;
2407         cursor.key_beg.obj_type = 0;
2408         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2409         cursor.key_beg.key = HAMMER_MIN_KEY;
2410
2411         cursor.key_end = cursor.key_beg;
2412         cursor.key_end.rec_type = 0xFFFF;
2413         cursor.key_end.key = HAMMER_MAX_KEY;
2414
2415         cursor.asof = ip->obj_asof;
2416         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2417
2418         error = hammer_ip_first(&cursor);
2419         if (error == ENOENT)
2420                 error = 0;
2421         else if (error == 0)
2422                 error = ENOTEMPTY;
2423         hammer_done_cursor(&cursor);
2424         return(error);
2425 }
2426