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 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
  48                                 hammer_btree_leaf_elm_t leaf);
  49
  50 struct rec_trunc_info {
  51         u_int16_t       rec_type;
  52         int64_t         trunc_off;
  53 };
  54
  55 struct hammer_bulk_info {
  56         hammer_record_t record;
  57         hammer_record_t conflict;
  58 };
  59
  60 /*
  61  * Red-black tree support.  Comparison code for insertion.
  62  */
  63 static int
  64 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  65 {
  66         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  67                 return(-1);
  68         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  69                 return(1);
  70
  71         if (rec1->leaf.base.key < rec2->leaf.base.key)
  72                 return(-1);
  73         if (rec1->leaf.base.key > rec2->leaf.base.key)
  74                 return(1);
  75
  76         /*
  77          * For search & insertion purposes records deleted by the
  78          * frontend or deleted/committed by the backend are silently
  79          * ignored.  Otherwise pipelined insertions will get messed
  80          * up.
  81          *
  82          * rec1 is greater then rec2 if rec1 is marked deleted.
  83          * rec1 is less then rec2 if rec2 is marked deleted.
  84          *
  85          * Multiple deleted records may be present, do not return 0
  86          * if both are marked deleted.
  87          */
  88         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  89                            HAMMER_RECF_COMMITTED)) {
  90                 return(1);
  91         }
  92         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  93                            HAMMER_RECF_COMMITTED)) {
  94                 return(-1);
  95         }
  96
  97         return(0);
  98 }
  99
 100 /*
 101  * Basic record comparison code similar to hammer_btree_cmp().
 102  *
 103  * obj_id is not compared and may not yet be assigned in the record.
 104  */
 105 static int
 106 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 107 {
 108         if (elm->rec_type < rec->leaf.base.rec_type)
 109                 return(-3);
 110         if (elm->rec_type > rec->leaf.base.rec_type)
 111                 return(3);
 112
 113         if (elm->key < rec->leaf.base.key)
 114                 return(-2);
 115         if (elm->key > rec->leaf.base.key)
 116                 return(2);
 117
 118         /*
 119          * Never match against an item deleted by the frontend
 120          * or backend, or committed by the backend.
 121          *
 122          * elm is less then rec if rec is marked deleted.
 123          */
 124         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 125                           HAMMER_RECF_COMMITTED)) {
 126                 return(-1);
 127         }
 128         return(0);
 129 }
 130
 131 /*
 132  * Ranged scan to locate overlapping record(s).  This is used by
 133  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 134  * to use a ranged scan because the keys for data records with the
 135  * same file base offset can be different due to differing data_len's.
 136  *
 137  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 138  */
 139 static int
 140 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 141 {
 142         struct hammer_bulk_info *info = data;
 143         hammer_btree_leaf_elm_t leaf = &info->record->leaf;
 144
 145         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 146                 return(-3);
 147         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 148                 return(3);
 149
 150         /*
 151          * Overlap compare
 152          */
 153         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 154                 /* rec_beg >= leaf_end */
 155                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 156                         return(2);
 157                 /* rec_end <= leaf_beg */
 158                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 159                         return(-2);
 160         } else {
 161                 if (rec->leaf.base.key < leaf->base.key)
 162                         return(-2);
 163                 if (rec->leaf.base.key > leaf->base.key)
 164                         return(2);
 165         }
 166
 167         /*
 168          * We have to return 0 at this point, even if DELETED_FE is set,
 169          * because returning anything else will cause the scan to ignore
 170          * one of the branches when we really want it to check both.
 171          */
 172         return(0);
 173 }
 174
 175 /*
 176  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 177  * is reversed so the comparison result has to be negated.  key_beg and
 178  * key_end are both range-inclusive.
 179  *
 180  * Localized deletions are not cached in-memory.
 181  */
 182 static
 183 int
 184 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 185 {
 186         hammer_cursor_t cursor = data;
 187         int r;
 188
 189         r = hammer_rec_cmp(&cursor->key_beg, rec);
 190         if (r > 1)
 191                 return(-1);
 192         r = hammer_rec_cmp(&cursor->key_end, rec);
 193         if (r < -1)
 194                 return(1);
 195         return(0);
 196 }
 197
 198 /*
 199  * This compare function is used when simply looking up key_beg.
 200  */
 201 static
 202 int
 203 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 204 {
 205         hammer_cursor_t cursor = data;
 206         int r;
 207
 208         r = hammer_rec_cmp(&cursor->key_beg, rec);
 209         if (r > 1)
 210                 return(-1);
 211         if (r < -1)
 212                 return(1);
 213         return(0);
 214 }
 215
 216 /*
 217  * Locate blocks within the truncation range.  Partial blocks do not count.
 218  */
 219 static
 220 int
 221 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 222 {
 223         struct rec_trunc_info *info = data;
 224
 225         if (rec->leaf.base.rec_type < info->rec_type)
 226                 return(-1);
 227         if (rec->leaf.base.rec_type > info->rec_type)
 228                 return(1);
 229
 230         switch(rec->leaf.base.rec_type) {
 231         case HAMMER_RECTYPE_DB:
 232                 /*
 233                  * DB record key is not beyond the truncation point, retain.
 234                  */
 235                 if (rec->leaf.base.key < info->trunc_off)
 236                         return(-1);
 237                 break;
 238         case HAMMER_RECTYPE_DATA:
 239                 /*
 240                  * DATA record offset start is not beyond the truncation point,
 241                  * retain.
 242                  */
 243                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 244                         return(-1);
 245                 break;
 246         default:
 247                 panic("hammer_rec_trunc_cmp: unexpected record type");
 248         }
 249
 250         /*
 251          * The record start is >= the truncation point, return match,
 252          * the record should be destroyed.
 253          */
 254         return(0);
 255 }
 256
 257 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 258
 259 /*
 260  * Allocate a record for the caller to finish filling in.  The record is
 261  * returned referenced.
 262  */
 263 hammer_record_t
 264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 265 {
 266         hammer_record_t record;
 267         hammer_mount_t hmp;
 268
 269         hmp = ip->hmp;
 270         ++hammer_count_records;
 271         record = kmalloc(sizeof(*record), hmp->m_misc,
 272                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 273         record->flush_state = HAMMER_FST_IDLE;
 274         record->ip = ip;
 275         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 276         record->leaf.data_len = data_len;
 277         hammer_ref(&record->lock);
 278
 279         if (data_len) {
 280                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
 281                 record->flags |= HAMMER_RECF_ALLOCDATA;
 282                 ++hammer_count_record_datas;
 283         }
 284
 285         return (record);
 286 }
 287
 288 void
 289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 290 {
 291         while (record->flush_state == HAMMER_FST_FLUSH) {
 292                 record->flags |= HAMMER_RECF_WANTED;
 293                 tsleep(record, 0, ident, 0);
 294         }
 295 }
 296
 297 /*
 298  * Called from the backend, hammer_inode.c, after a record has been
 299  * flushed to disk.  The record has been exclusively locked by the
 300  * caller and interlocked with BE.
 301  *
 302  * We clean up the state, unlock, and release the record (the record
 303  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 304  */
 305 void
 306 hammer_flush_record_done(hammer_record_t record, int error)
 307 {
 308         hammer_inode_t target_ip;
 309
 310         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 311         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 312
 313         /*
 314          * If an error occured, the backend was unable to sync the
 315          * record to its media.  Leave the record intact.
 316          */
 317         if (error) {
 318                 hammer_critical_error(record->ip->hmp, record->ip, error,
 319                                       "while flushing record");
 320         }
 321
 322         --record->flush_group->refs;
 323         record->flush_group = NULL;
 324
 325         /*
 326          * Adjust the flush state and dependancy based on success or
 327          * failure.
 328          */
 329         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
 330                 if ((target_ip = record->target_ip) != NULL) {
 331                         TAILQ_REMOVE(&target_ip->target_list, record,
 332                                      target_entry);
 333                         record->target_ip = NULL;
 334                         hammer_test_inode(target_ip);
 335                 }
 336                 record->flush_state = HAMMER_FST_IDLE;
 337         } else {
 338                 if (record->target_ip) {
 339                         record->flush_state = HAMMER_FST_SETUP;
 340                         hammer_test_inode(record->ip);
 341                         hammer_test_inode(record->target_ip);
 342                 } else {
 343                         record->flush_state = HAMMER_FST_IDLE;
 344                 }
 345         }
 346         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 347
 348         /*
 349          * Cleanup
 350          */
 351         if (record->flags & HAMMER_RECF_WANTED) {
 352                 record->flags &= ~HAMMER_RECF_WANTED;
 353                 wakeup(record);
 354         }
 355         hammer_rel_mem_record(record);
 356 }
 357
 358 /*
 359  * Release a memory record.  Records marked for deletion are immediately
 360  * removed from the RB-Tree but otherwise left intact until the last ref
 361  * goes away.
 362  */
 363 void
 364 hammer_rel_mem_record(struct hammer_record *record)
 365 {
 366         hammer_mount_t hmp;
 367         hammer_reserve_t resv;
 368         hammer_inode_t ip;
 369         hammer_inode_t target_ip;
 370         int diddrop;
 371
 372         hammer_rel(&record->lock);
 373
 374         if (hammer_norefs(&record->lock)) {
 375                 /*
 376                  * Upon release of the last reference wakeup any waiters.
 377                  * The record structure may get destroyed so callers will
 378                  * loop up and do a relookup.
 379                  *
 380                  * WARNING!  Record must be removed from RB-TREE before we
 381                  * might possibly block.  hammer_test_inode() can block!
 382                  */
 383                 ip = record->ip;
 384                 hmp = ip->hmp;
 385
 386                 /*
 387                  * Upon release of the last reference a record marked deleted
 388                  * by the front or backend, or committed by the backend,
 389                  * is destroyed.
 390                  */
 391                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 392                                      HAMMER_RECF_DELETED_BE |
 393                                      HAMMER_RECF_COMMITTED)) {
 394                         KKASSERT(hammer_isactive(&ip->lock) > 0);
 395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 396
 397                         /*
 398                          * target_ip may have zero refs, we have to ref it
 399                          * to prevent it from being ripped out from under
 400                          * us.
 401                          */
 402                         if ((target_ip = record->target_ip) != NULL) {
 403                                 TAILQ_REMOVE(&target_ip->target_list,
 404                                              record, target_entry);
 405                                 record->target_ip = NULL;
 406                                 hammer_ref(&target_ip->lock);
 407                         }
 408
 409                         /*
 410                          * Remove the record from the B-Tree
 411                          */
 412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 413                                 RB_REMOVE(hammer_rec_rb_tree,
 414                                           &record->ip->rec_tree,
 415                                           record);
 416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 417                                 KKASSERT(ip->rsv_recs > 0);
 418                                 diddrop = 1;
 419                         } else {
 420                                 diddrop = 0;
 421                         }
 422
 423                         /*
 424                          * We must wait for any direct-IO to complete before
 425                          * we can destroy the record because the bio may
 426                          * have a reference to it.
 427                          */
 428                         if (record->gflags &
 429                            (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
 430                                 hammer_io_direct_wait(record);
 431                         }
 432
 433                         /*
 434                          * Account for the completion after the direct IO
 435                          * has completed.
 436                          */
 437                         if (diddrop) {
 438                                 --hmp->rsv_recs;
 439                                 --ip->rsv_recs;
 440                                 hmp->rsv_databytes -= record->leaf.data_len;
 441
 442                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 443                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 444                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 445                                         hammer_test_inode(record->ip);
 446                                 }
 447                                 if (ip->rsv_recs == hammer_limit_inode_recs - 1)
 448                                         wakeup(&ip->rsv_recs);
 449                         }
 450
 451                         /*
 452                          * Do this test after removing record from the B-Tree.
 453                          */
 454                         if (target_ip) {
 455                                 hammer_test_inode(target_ip);
 456                                 hammer_rel_inode(target_ip, 0);
 457                         }
 458
 459                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 460                                 --hammer_count_record_datas;
 461                                 kfree(record->data, hmp->m_misc);
 462                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 463                         }
 464
 465                         /*
 466                          * Release the reservation.
 467                          *
 468                          * If the record was not committed we can theoretically
 469                          * undo the reservation.  However, doing so might
 470                          * create weird edge cases with the ordering of
 471                          * direct writes because the related buffer cache
 472                          * elements are per-vnode.  So we don't try.
 473                          */
 474                         if ((resv = record->resv) != NULL) {
 475                                 /* XXX undo leaf.data_offset,leaf.data_len */
 476                                 hammer_blockmap_reserve_complete(hmp, resv);
 477                                 record->resv = NULL;
 478                         }
 479                         record->data = NULL;
 480                         --hammer_count_records;
 481                         kfree(record, hmp->m_misc);
 482                 }
 483         }
 484 }
 485
 486 /*
 487  * Record visibility depends on whether the record is being accessed by
 488  * the backend or the frontend.  Backend tests ignore the frontend delete
 489  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 490  * must also check for commit races.
 491  *
 492  * Return non-zero if the record is visible, zero if it isn't or if it is
 493  * deleted.  Returns 0 if the record has been comitted (unless the special
 494  * delete-visibility flag is set).  A committed record must be located
 495  * via the media B-Tree.  Returns non-zero if the record is good.
 496  *
 497  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 498  * records to be returned.  This is so pending deletions are detected
 499  * when using an iterator to locate an unused hash key, or when we need
 500  * to locate historical records on-disk to destroy.
 501  */
 502 static __inline
 503 int
 504 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 505 {
 506         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 507                 return(1);
 508         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 509                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 510                                      HAMMER_RECF_COMMITTED)) {
 511                         return(0);
 512                 }
 513         } else {
 514                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 515                                      HAMMER_RECF_DELETED_BE |
 516                                      HAMMER_RECF_COMMITTED)) {
 517                         return(0);
 518                 }
 519         }
 520         return(1);
 521 }
 522
 523 /*
 524  * This callback is used as part of the RB_SCAN function for in-memory
 525  * records.  We terminate it (return -1) as soon as we get a match.
 526  *
 527  * This routine is used by frontend code.
 528  *
 529  * The primary compare code does not account for ASOF lookups.  This
 530  * code handles that case as well as a few others.
 531  */
 532 static
 533 int
 534 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 535 {
 536         hammer_cursor_t cursor = data;
 537
 538         /*
 539          * We terminate on success, so this should be NULL on entry.
 540          */
 541         KKASSERT(cursor->iprec == NULL);
 542
 543         /*
 544          * Skip if the record was marked deleted or committed.
 545          */
 546         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 547                 return(0);
 548
 549         /*
 550          * Skip if not visible due to our as-of TID
 551          */
 552         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 553                 if (cursor->asof < rec->leaf.base.create_tid)
 554                         return(0);
 555                 if (rec->leaf.base.delete_tid &&
 556                     cursor->asof >= rec->leaf.base.delete_tid) {
 557                         return(0);
 558                 }
 559         }
 560
 561         /*
 562          * ref the record.  The record is protected from backend B-Tree
 563          * interactions by virtue of the cursor's IP lock.
 564          */
 565         hammer_ref(&rec->lock);
 566
 567         /*
 568          * The record may have been deleted or committed while we
 569          * were blocked.  XXX remove?
 570          */
 571         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 572                 hammer_rel_mem_record(rec);
 573                 return(0);
 574         }
 575
 576         /*
 577          * Set the matching record and stop the scan.
 578          */
 579         cursor->iprec = rec;
 580         return(-1);
 581 }
 582
 583
 584 /*
 585  * Lookup an in-memory record given the key specified in the cursor.  Works
 586  * just like hammer_btree_lookup() but operates on an inode's in-memory
 587  * record list.
 588  *
 589  * The lookup must fail if the record is marked for deferred deletion.
 590  *
 591  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 592  */
 593 static
 594 int
 595 hammer_mem_lookup(hammer_cursor_t cursor)
 596 {
 597         KKASSERT(cursor->ip);
 598         if (cursor->iprec) {
 599                 hammer_rel_mem_record(cursor->iprec);
 600                 cursor->iprec = NULL;
 601         }
 602         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 603                                    hammer_rec_scan_callback, cursor);
 604
 605         return (cursor->iprec ? 0 : ENOENT);
 606 }
 607
 608 /*
 609  * hammer_mem_first() - locate the first in-memory record matching the
 610  * cursor within the bounds of the key range.
 611  *
 612  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 613  * will set ATEMEM the same as MEMEOF, and does not return any error.
 614  */
 615 static
 616 void
 617 hammer_mem_first(hammer_cursor_t cursor)
 618 {
 619         hammer_inode_t ip;
 620
 621         ip = cursor->ip;
 622         KKASSERT(ip != NULL);
 623
 624         if (cursor->iprec) {
 625                 hammer_rel_mem_record(cursor->iprec);
 626                 cursor->iprec = NULL;
 627         }
 628         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 629                                    hammer_rec_scan_callback, cursor);
 630
 631         if (cursor->iprec)
 632                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
 633         else
 634                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
 635 }
 636
 637 /************************************************************************
 638  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 639  ************************************************************************
 640  *
 641  * These functions manipulate in-memory records.  Such records typically
 642  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 643  */
 644
 645 /*
 646  * Add a directory entry (dip,ncp) which references inode (ip).
 647  *
 648  * Note that the low 32 bits of the namekey are set temporarily to create
 649  * a unique in-memory record, and may be modified a second time when the
 650  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 651  * all 0's when synching to disk, which is not handled here.
 652  *
 653  * NOTE: bytes does not include any terminating \0 on name, and name might
 654  * not be terminated.
 655  */
 656 int
 657 hammer_ip_add_directory(struct hammer_transaction *trans,
 658                      struct hammer_inode *dip, const char *name, int bytes,
 659                      struct hammer_inode *ip)
 660 {
 661         struct hammer_cursor cursor;
 662         hammer_record_t record;
 663         int error;
 664         u_int32_t max_iterations;
 665
 666         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 667
 668         record->type = HAMMER_MEM_RECORD_ADD;
 669         record->leaf.base.localization = dip->obj_localization +
 670                                          hammer_dir_localization(dip);
 671         record->leaf.base.obj_id = dip->obj_id;
 672         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 673                                                          &max_iterations);
 674         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 675         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 676         record->data->entry.obj_id = ip->obj_id;
 677         record->data->entry.localization = ip->obj_localization;
 678         bcopy(name, record->data->entry.name, bytes);
 679
 680         ++ip->ino_data.nlinks;
 681         ip->ino_data.ctime = trans->time;
 682         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 683
 684         /*
 685          * Find an unused namekey.  Both the in-memory record tree and
 686          * the B-Tree are checked.  We do not want historically deleted
 687          * names to create a collision as our iteration space may be limited,
 688          * and since create_tid wouldn't match anyway an ASOF search
 689          * must be used to locate collisions.
 690          *
 691          * delete-visibility is set so pending deletions do not give us
 692          * a false-negative on our ability to use an iterator.
 693          *
 694          * The iterator must not rollover the key.  Directory keys only
 695          * use the positive key space.
 696          */
 697         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 698         cursor.key_beg = record->leaf.base;
 699         cursor.flags |= HAMMER_CURSOR_ASOF;
 700         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 701         cursor.asof = ip->obj_asof;
 702
 703         while (hammer_ip_lookup(&cursor) == 0) {
 704                 ++record->leaf.base.key;
 705                 KKASSERT(record->leaf.base.key > 0);
 706                 cursor.key_beg.key = record->leaf.base.key;
 707                 if (--max_iterations == 0) {
 708                         hammer_rel_mem_record(record);
 709                         error = ENOSPC;
 710                         goto failed;
 711                 }
 712         }
 713
 714         /*
 715          * The target inode and the directory entry are bound together.
 716          */
 717         record->target_ip = ip;
 718         record->flush_state = HAMMER_FST_SETUP;
 719         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 720
 721         /*
 722          * The inode now has a dependancy and must be taken out of the idle
 723          * state.  An inode not in an idle state is given an extra reference.
 724          *
 725          * When transitioning to a SETUP state flag for an automatic reflush
 726          * when the dependancies are disposed of if someone is waiting on
 727          * the inode.
 728          */
 729         if (ip->flush_state == HAMMER_FST_IDLE) {
 730                 hammer_ref(&ip->lock);
 731                 ip->flush_state = HAMMER_FST_SETUP;
 732                 if (ip->flags & HAMMER_INODE_FLUSHW)
 733                         ip->flags |= HAMMER_INODE_REFLUSH;
 734         }
 735         error = hammer_mem_add(record);
 736         if (error == 0) {
 737                 dip->ino_data.mtime = trans->time;
 738                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 739         }
 740 failed:
 741         hammer_done_cursor(&cursor);
 742         return(error);
 743 }
 744
 745 /*
 746  * Delete the directory entry and update the inode link count.  The
 747  * cursor must be seeked to the directory entry record being deleted.
 748  *
 749  * The related inode should be share-locked by the caller.  The caller is
 750  * on the frontend.  It could also be NULL indicating that the directory
 751  * entry being removed has no related inode.
 752  *
 753  * This function can return EDEADLK requiring the caller to terminate
 754  * the cursor, any locks, wait on the returned record, and retry.
 755  */
 756 int
 757 hammer_ip_del_directory(struct hammer_transaction *trans,
 758                      hammer_cursor_t cursor, struct hammer_inode *dip,
 759                      struct hammer_inode *ip)
 760 {
 761         hammer_record_t record;
 762         int error;
 763
 764         if (hammer_cursor_inmem(cursor)) {
 765                 /*
 766                  * In-memory (unsynchronized) records can simply be freed.
 767                  *
 768                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 769                  * by the backend, we must still avoid races against the
 770                  * backend potentially syncing the record to the media.
 771                  *
 772                  * We cannot call hammer_ip_delete_record(), that routine may
 773                  * only be called from the backend.
 774                  */
 775                 record = cursor->iprec;
 776                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 777                                      HAMMER_RECF_DELETED_BE |
 778                                      HAMMER_RECF_COMMITTED)) {
 779                         KKASSERT(cursor->deadlk_rec == NULL);
 780                         hammer_ref(&record->lock);
 781                         cursor->deadlk_rec = record;
 782                         error = EDEADLK;
 783                 } else {
 784                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 785                         record->flags |= HAMMER_RECF_DELETED_FE;
 786                         error = 0;
 787                 }
 788         } else {
 789                 /*
 790                  * If the record is on-disk we have to queue the deletion by
 791                  * the record's key.  This also causes lookups to skip the
 792                  * record (lookups for the purposes of finding an unused
 793                  * directory key do not skip the record).
 794                  */
 795                 KKASSERT(dip->flags &
 796                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 797                 record = hammer_alloc_mem_record(dip, 0);
 798                 record->type = HAMMER_MEM_RECORD_DEL;
 799                 record->leaf.base = cursor->leaf->base;
 800                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
 801
 802                 /*
 803                  * ip may be NULL, indicating the deletion of a directory
 804                  * entry which has no related inode.
 805                  */
 806                 record->target_ip = ip;
 807                 if (ip) {
 808                         record->flush_state = HAMMER_FST_SETUP;
 809                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 810                                           target_entry);
 811                 } else {
 812                         record->flush_state = HAMMER_FST_IDLE;
 813                 }
 814
 815                 /*
 816                  * The inode now has a dependancy and must be taken out of
 817                  * the idle state.  An inode not in an idle state is given
 818                  * an extra reference.
 819                  *
 820                  * When transitioning to a SETUP state flag for an automatic
 821                  * reflush when the dependancies are disposed of if someone
 822                  * is waiting on the inode.
 823                  */
 824                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 825                         hammer_ref(&ip->lock);
 826                         ip->flush_state = HAMMER_FST_SETUP;
 827                         if (ip->flags & HAMMER_INODE_FLUSHW)
 828                                 ip->flags |= HAMMER_INODE_REFLUSH;
 829                 }
 830
 831                 error = hammer_mem_add(record);
 832         }
 833
 834         /*
 835          * One less link.  The file may still be open in the OS even after
 836          * all links have gone away.
 837          *
 838          * We have to terminate the cursor before syncing the inode to
 839          * avoid deadlocking against ourselves.  XXX this may no longer
 840          * be true.
 841          *
 842          * If nlinks drops to zero and the vnode is inactive (or there is
 843          * no vnode), call hammer_inode_unloadable_check() to zonk the
 844          * inode.  If we don't do this here the inode will not be destroyed
 845          * on-media until we unmount.
 846          */
 847         if (error == 0) {
 848                 if (ip) {
 849                         --ip->ino_data.nlinks;  /* do before we might block */
 850                         ip->ino_data.ctime = trans->time;
 851                 }
 852                 dip->ino_data.mtime = trans->time;
 853                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 854                 if (ip) {
 855                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 856                         if (ip->ino_data.nlinks == 0 &&
 857                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 858                                 hammer_done_cursor(cursor);
 859                                 hammer_inode_unloadable_check(ip, 1);
 860                                 hammer_flush_inode(ip, 0);
 861                         }
 862                 }
 863
 864         }
 865         return(error);
 866 }
 867
 868 /*
 869  * Add a record to an inode.
 870  *
 871  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 872  * initialize the following additional fields:
 873  *
 874  * The related inode should be share-locked by the caller.  The caller is
 875  * on the frontend.
 876  *
 877  * record->rec.entry.base.base.key
 878  * record->rec.entry.base.base.rec_type
 879  * record->rec.entry.base.base.data_len
 880  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 881  */
 882 int
 883 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 884 {
 885         hammer_inode_t ip = record->ip;
 886         int error;
 887
 888         KKASSERT(record->leaf.base.localization != 0);
 889         record->leaf.base.obj_id = ip->obj_id;
 890         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 891         error = hammer_mem_add(record);
 892         return(error);
 893 }
 894
 895 /*
 896  * Locate a pre-existing bulk record in memory.  The caller wishes to
 897  * replace the record with a new one.  The existing record may have a
 898  * different length (and thus a different key) so we have to use an
 899  * overlap check function.
 900  */
 901 static hammer_record_t
 902 hammer_ip_get_bulk(hammer_record_t record)
 903 {
 904         struct hammer_bulk_info info;
 905         hammer_inode_t ip = record->ip;
 906
 907         info.record = record;
 908         info.conflict = NULL;
 909         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 910                                    hammer_bulk_scan_callback, &info);
 911
 912         return(info.conflict);  /* may be NULL */
 913 }
 914
 915 /*
 916  * Take records vetted by overlap_cmp.  The first non-deleted record
 917  * (if any) stops the scan.
 918  */
 919 static int
 920 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 921 {
 922         struct hammer_bulk_info *info = data;
 923
 924         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 925                              HAMMER_RECF_COMMITTED)) {
 926                 return(0);
 927         }
 928         hammer_ref(&record->lock);
 929         info->conflict = record;
 930         return(-1);                     /* stop scan */
 931 }
 932
 933 /*
 934  * Reserve blockmap space placemarked with an in-memory record.
 935  *
 936  * This routine is called by the frontend in order to be able to directly
 937  * flush a buffer cache buffer.  The frontend has locked the related buffer
 938  * cache buffers and we should be able to manipulate any overlapping
 939  * in-memory records.
 940  *
 941  * The caller is responsible for adding the returned record and deleting
 942  * the returned conflicting record (if any), typically by calling
 943  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
 944  */
 945 hammer_record_t
 946 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 947                    int *errorp)
 948 {
 949         hammer_record_t record;
 950         int zone;
 951
 952         /*
 953          * Create a record to cover the direct write.  The record cannot
 954          * be added to the in-memory RB tree here as it might conflict
 955          * with an existing memory record.  See hammer_io_direct_write().
 956          *
 957          * The backend is responsible for finalizing the space reserved in
 958          * this record.
 959          *
 960          * XXX bytes not aligned, depend on the reservation code to
 961          * align the reservation.
 962          */
 963         record = hammer_alloc_mem_record(ip, 0);
 964         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 965                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 966         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 967                                                &record->leaf.data_offset,
 968                                                errorp);
 969         if (record->resv == NULL) {
 970                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 971                 hammer_rel_mem_record(record);
 972                 return(NULL);
 973         }
 974         record->type = HAMMER_MEM_RECORD_DATA;
 975         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 976         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 977         record->leaf.base.obj_id = ip->obj_id;
 978         record->leaf.base.key = file_offset + bytes;
 979         record->leaf.base.localization = ip->obj_localization +
 980                                          HAMMER_LOCALIZE_MISC;
 981         record->leaf.data_len = bytes;
 982         hammer_crc_set_leaf(data, &record->leaf);
 983         KKASSERT(*errorp == 0);
 984
 985         return(record);
 986 }
 987
 988 /*
 989  * Called by hammer_io_direct_write() prior to any possible completion
 990  * of the BIO to emplace the memory record associated with the I/O and
 991  * to replace any prior memory record which might still be active.
 992  *
 993  * Setting the FE deleted flag on the old record (if any) avoids any RB
 994  * tree insertion conflict, amoung other things.
 995  *
 996  * This has to be done prior to the caller completing any related buffer
 997  * cache I/O or a reinstantiation of the buffer may load data from the
 998  * old media location instead of the new media location.  The holding
 999  * of the locked buffer cache buffer serves to interlock the record
1000  * replacement operation.
1001  */
1002 void
1003 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1004 {
1005         hammer_record_t conflict;
1006         int error;
1007
1008         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1009                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1010                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1011                         break;
1012                 }
1013                 conflict->flags |= HAMMER_RECF_WANTED;
1014                 tsleep(conflict, 0, "hmrrc3", 0);
1015                 hammer_rel_mem_record(conflict);
1016         }
1017         error = hammer_mem_add(record);
1018         if (conflict)
1019                 hammer_rel_mem_record(conflict);
1020         KKASSERT(error == 0);
1021 }
1022
1023 /*
1024  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1025  * and records in a flushing state are handled by the backend.  The vnops
1026  * setattr code will handle the block containing the truncation point.
1027  *
1028  * Partial blocks are not deleted.
1029  *
1030  * This code is only called on regular files.
1031  */
1032 int
1033 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1034 {
1035         struct rec_trunc_info info;
1036
1037         switch(ip->ino_data.obj_type) {
1038         case HAMMER_OBJTYPE_REGFILE:
1039                 info.rec_type = HAMMER_RECTYPE_DATA;
1040                 break;
1041         case HAMMER_OBJTYPE_DBFILE:
1042                 info.rec_type = HAMMER_RECTYPE_DB;
1043                 break;
1044         default:
1045                 return(EINVAL);
1046         }
1047         info.trunc_off = file_size;
1048         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1049                                    hammer_frontend_trunc_callback, &info);
1050         return(0);
1051 }
1052
1053 /*
1054  * Scan callback for frontend records to destroy during a truncation.
1055  * We must ensure that DELETED_FE is set on the record or the frontend
1056  * will get confused in future read() calls.
1057  *
1058  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1059  *       In this rare case we must wait for the interlock to be cleared.
1060  *
1061  * NOTE: This function is only called on regular files.  There are further
1062  *       restrictions to the setting of DELETED_FE on directory records
1063  *       undergoing a flush due to sensitive inode link count calculations.
1064  */
1065 static int
1066 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1067 {
1068         if (record->flags & HAMMER_RECF_DELETED_FE)
1069                 return(0);
1070 #if 0
1071         if (record->flush_state == HAMMER_FST_FLUSH)
1072                 return(0);
1073 #endif
1074         hammer_ref(&record->lock);
1075         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1076                 hammer_wait_mem_record_ident(record, "hmmtrr");
1077         record->flags |= HAMMER_RECF_DELETED_FE;
1078         hammer_rel_mem_record(record);
1079         return(0);
1080 }
1081
1082 /*
1083  * Return 1 if the caller must check for and delete existing records
1084  * before writing out a new data record.
1085  *
1086  * Return 0 if the caller can just insert the record into the B-Tree without
1087  * checking.
1088  */
1089 static int
1090 hammer_record_needs_overwrite_delete(hammer_record_t record)
1091 {
1092         hammer_inode_t ip = record->ip;
1093         int64_t file_offset;
1094         int r;
1095
1096         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1097                 file_offset = record->leaf.base.key;
1098         else
1099                 file_offset = record->leaf.base.key - record->leaf.data_len;
1100         r = (file_offset < ip->save_trunc_off);
1101         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1102                 if (ip->save_trunc_off <= record->leaf.base.key)
1103                         ip->save_trunc_off = record->leaf.base.key + 1;
1104         } else {
1105                 if (ip->save_trunc_off < record->leaf.base.key)
1106                         ip->save_trunc_off = record->leaf.base.key;
1107         }
1108         return(r);
1109 }
1110
1111 /*
1112  * Backend code.  Sync a record to the media.
1113  */
1114 int
1115 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1116 {
1117         hammer_transaction_t trans = cursor->trans;
1118         int64_t file_offset;
1119         int bytes;
1120         void *bdata;
1121         int error;
1122         int doprop;
1123
1124         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1125         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1126         KKASSERT(record->leaf.base.localization != 0);
1127
1128         /*
1129          * Any direct-write related to the record must complete before we
1130          * can sync the record to the on-disk media.
1131          */
1132         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1133                 hammer_io_direct_wait(record);
1134
1135         /*
1136          * If this is a bulk-data record placemarker there may be an existing
1137          * record on-disk, indicating a data overwrite.  If there is the
1138          * on-disk record must be deleted before we can insert our new record.
1139          *
1140          * We've synthesized this record and do not know what the create_tid
1141          * on-disk is, nor how much data it represents.
1142          *
1143          * Keep in mind that (key) for data records is (base_offset + len),
1144          * not (base_offset).  Also, we only want to get rid of on-disk
1145          * records since we are trying to sync our in-memory record, call
1146          * hammer_ip_delete_range() with truncating set to 1 to make sure
1147          * it skips in-memory records.
1148          *
1149          * It is ok for the lookup to return ENOENT.
1150          *
1151          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1152          * to call hammer_ip_delete_range() or not.  This also means we must
1153          * update sync_trunc_off() as we write.
1154          */
1155         if (record->type == HAMMER_MEM_RECORD_DATA &&
1156             hammer_record_needs_overwrite_delete(record)) {
1157                 file_offset = record->leaf.base.key - record->leaf.data_len;
1158                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1159                         ~HAMMER_BUFMASK;
1160                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1161                 error = hammer_ip_delete_range(
1162                                 cursor, record->ip,
1163                                 file_offset, file_offset + bytes - 1,
1164                                 1);
1165                 if (error && error != ENOENT)
1166                         goto done;
1167         }
1168
1169         /*
1170          * If this is a general record there may be an on-disk version
1171          * that must be deleted before we can insert the new record.
1172          */
1173         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1174                 error = hammer_delete_general(cursor, record->ip,
1175                                               &record->leaf);
1176                 if (error && error != ENOENT)
1177                         goto done;
1178         }
1179
1180         /*
1181          * Setup the cursor.
1182          */
1183         hammer_normalize_cursor(cursor);
1184         cursor->key_beg = record->leaf.base;
1185         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1186         cursor->flags |= HAMMER_CURSOR_BACKEND;
1187         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1188
1189         /*
1190          * Records can wind up on-media before the inode itself is on-media.
1191          * Flag the case.
1192          */
1193         record->ip->flags |= HAMMER_INODE_DONDISK;
1194
1195         /*
1196          * If we are deleting a directory entry an exact match must be
1197          * found on-disk.
1198          */
1199         if (record->type == HAMMER_MEM_RECORD_DEL) {
1200                 error = hammer_btree_lookup(cursor);
1201                 if (error == 0) {
1202                         KKASSERT(cursor->iprec == NULL);
1203                         error = hammer_ip_delete_record(cursor, record->ip,
1204                                                         trans->tid);
1205                         if (error == 0) {
1206                                 record->flags |= HAMMER_RECF_DELETED_BE |
1207                                                  HAMMER_RECF_COMMITTED;
1208                                 ++record->ip->rec_generation;
1209                         }
1210                 }
1211                 goto done;
1212         }
1213
1214         /*
1215          * We are inserting.
1216          *
1217          * Issue a lookup to position the cursor and locate the insertion
1218          * point.  The target key should not exist.  If we are creating a
1219          * directory entry we may have to iterate the low 32 bits of the
1220          * key to find an unused key.
1221          */
1222         hammer_sync_lock_sh(trans);
1223         cursor->flags |= HAMMER_CURSOR_INSERT;
1224         error = hammer_btree_lookup(cursor);
1225         if (hammer_debug_inode)
1226                 kprintf("DOINSERT LOOKUP %d\n", error);
1227         if (error == 0) {
1228                 kprintf("hammer_ip_sync_record: duplicate rec "
1229                         "at (%016llx)\n", (long long)record->leaf.base.key);
1230                 if (hammer_debug_critical)
1231                         Debugger("duplicate record1");
1232                 error = EIO;
1233         }
1234 #if 0
1235         if (record->type == HAMMER_MEM_RECORD_DATA)
1236                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1237                         record->leaf.base.key - record->leaf.data_len,
1238                         record->leaf.data_offset, error);
1239 #endif
1240
1241         if (error != ENOENT)
1242                 goto done_unlock;
1243
1244         /*
1245          * Allocate the record and data.  The result buffers will be
1246          * marked as being modified and further calls to
1247          * hammer_modify_buffer() will result in unneeded UNDO records.
1248          *
1249          * Support zero-fill records (data == NULL and data_len != 0)
1250          */
1251         if (record->type == HAMMER_MEM_RECORD_DATA) {
1252                 /*
1253                  * The data portion of a bulk-data record has already been
1254                  * committed to disk, we need only adjust the layer2
1255                  * statistics in the same transaction as our B-Tree insert.
1256                  */
1257                 KKASSERT(record->leaf.data_offset != 0);
1258                 error = hammer_blockmap_finalize(trans,
1259                                                  record->resv,
1260                                                  record->leaf.data_offset,
1261                                                  record->leaf.data_len);
1262         } else if (record->data && record->leaf.data_len) {
1263                 /*
1264                  * Wholely cached record, with data.  Allocate the data.
1265                  */
1266                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1267                                           record->leaf.base.rec_type,
1268                                           &record->leaf.data_offset,
1269                                           &cursor->data_buffer,
1270                                           0, &error);
1271                 if (bdata == NULL)
1272                         goto done_unlock;
1273                 hammer_crc_set_leaf(record->data, &record->leaf);
1274                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1275                 bcopy(record->data, bdata, record->leaf.data_len);
1276                 hammer_modify_buffer_done(cursor->data_buffer);
1277         } else {
1278                 /*
1279                  * Wholely cached record, without data.
1280                  */
1281                 record->leaf.data_offset = 0;
1282                 record->leaf.data_crc = 0;
1283         }
1284
1285         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1286         if (hammer_debug_inode && error) {
1287                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1288                         error,
1289                         (long long)cursor->node->node_offset,
1290                         cursor->index,
1291                         (long long)record->leaf.base.key);
1292         }
1293
1294         /*
1295          * Our record is on-disk and we normally mark the in-memory version
1296          * as having been committed (and not BE-deleted).
1297          *
1298          * If the record represented a directory deletion but we had to
1299          * sync a valid directory entry to disk due to dependancies,
1300          * we must convert the record to a covering delete so the
1301          * frontend does not have visibility on the synced entry.
1302          *
1303          * WARNING: cursor's leaf pointer may have changed after do_propagation
1304          *          returns!
1305          */
1306         if (error == 0) {
1307                 if (doprop) {
1308                         hammer_btree_do_propagation(cursor,
1309                                                     record->ip->pfsm,
1310                                                     &record->leaf);
1311                 }
1312                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1313                         /*
1314                          * Must convert deleted directory entry add
1315                          * to a directory entry delete.
1316                          */
1317                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1318                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1319                         record->type = HAMMER_MEM_RECORD_DEL;
1320                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1321                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1322                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1323                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1324                                                  HAMMER_RECF_DELETED_BE)) == 0);
1325                         /* converted record is not yet committed */
1326                         /* hammer_flush_record_done takes care of the rest */
1327                 } else {
1328                         /*
1329                          * Everything went fine and we are now done with
1330                          * this record.
1331                          */
1332                         record->flags |= HAMMER_RECF_COMMITTED;
1333                         ++record->ip->rec_generation;
1334                 }
1335         } else {
1336                 if (record->leaf.data_offset) {
1337                         hammer_blockmap_free(trans, record->leaf.data_offset,
1338                                              record->leaf.data_len);
1339                 }
1340         }
1341 done_unlock:
1342         hammer_sync_unlock(trans);
1343 done:
1344         return(error);
1345 }
1346
1347 /*
1348  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1349  * entry's key is used to deal with hash collisions in the upper 32 bits.
1350  * A unique 64 bit key is generated in-memory and may be regenerated a
1351  * second time when the directory record is flushed to the on-disk B-Tree.
1352  *
1353  * A referenced record is passed to this function.  This function
1354  * eats the reference.  If an error occurs the record will be deleted.
1355  *
1356  * A copy of the temporary record->data pointer provided by the caller
1357  * will be made.
1358  */
1359 int
1360 hammer_mem_add(hammer_record_t record)
1361 {
1362         hammer_mount_t hmp = record->ip->hmp;
1363
1364         /*
1365          * Make a private copy of record->data
1366          */
1367         if (record->data)
1368                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1369
1370         /*
1371          * Insert into the RB tree.  A unique key should have already
1372          * been selected if this is a directory entry.
1373          */
1374         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1375                 record->flags |= HAMMER_RECF_DELETED_FE;
1376                 hammer_rel_mem_record(record);
1377                 return (EEXIST);
1378         }
1379         ++hmp->count_newrecords;
1380         ++hmp->rsv_recs;
1381         ++record->ip->rsv_recs;
1382         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1383         record->flags |= HAMMER_RECF_ONRBTREE;
1384         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1385         hammer_rel_mem_record(record);
1386         return(0);
1387 }
1388
1389 /************************************************************************
1390  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1391  ************************************************************************
1392  *
1393  * These functions augment the B-Tree scanning functions in hammer_btree.c
1394  * by merging in-memory records with on-disk records.
1395  */
1396
1397 /*
1398  * Locate a particular record either in-memory or on-disk.
1399  *
1400  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1401  * NOT be called to iterate results.
1402  */
1403 int
1404 hammer_ip_lookup(hammer_cursor_t cursor)
1405 {
1406         int error;
1407
1408         /*
1409          * If the element is in-memory return it without searching the
1410          * on-disk B-Tree
1411          */
1412         KKASSERT(cursor->ip);
1413         error = hammer_mem_lookup(cursor);
1414         if (error == 0) {
1415                 cursor->leaf = &cursor->iprec->leaf;
1416                 return(error);
1417         }
1418         if (error != ENOENT)
1419                 return(error);
1420
1421         /*
1422          * If the inode has on-disk components search the on-disk B-Tree.
1423          */
1424         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1425                 return(error);
1426         error = hammer_btree_lookup(cursor);
1427         if (error == 0)
1428                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1429         return(error);
1430 }
1431
1432 /*
1433  * Helper for hammer_ip_first()/hammer_ip_next()
1434  *
1435  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1436  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1437  * state if hammer_ip_next() needs to re-seek.
1438  */
1439 static __inline
1440 int
1441 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1442 {
1443         hammer_inode_t ip = cursor->ip;
1444         int error;
1445
1446         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1447                 error = hammer_btree_lookup(cursor);
1448                 if (error == ENOENT || error == EDEADLK) {
1449                         if (hammer_debug_general & 0x2000) {
1450                                 kprintf("error %d node %p %016llx index %d\n",
1451                                         error, cursor->node,
1452                                         (long long)cursor->node->node_offset,
1453                                         cursor->index);
1454                         }
1455                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1456                         error = hammer_btree_iterate(cursor);
1457                 }
1458                 if (error == 0) {
1459                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1460                                            HAMMER_CURSOR_ATEDISK);
1461                 } else {
1462                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1463                                          HAMMER_CURSOR_ATEDISK;
1464                         if (error == ENOENT)
1465                                 error = 0;
1466                 }
1467         } else {
1468                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1469                 error = 0;
1470         }
1471         return(error);
1472 }
1473
1474 /*
1475  * Helper for hammer_ip_next()
1476  *
1477  * The caller has determined that the media cursor is further along than the
1478  * memory cursor and must be reseeked after a generation number change.
1479  */
1480 static
1481 int
1482 _hammer_ip_reseek(hammer_cursor_t cursor)
1483 {
1484         struct hammer_base_elm save;
1485         hammer_btree_elm_t elm;
1486         int error;
1487         int r;
1488         int again = 0;
1489
1490         /*
1491          * Do the re-seek.
1492          */
1493         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1494                 (long long)cursor->ip->obj_id);
1495         save = cursor->key_beg;
1496         cursor->key_beg = cursor->iprec->leaf.base;
1497         error = _hammer_ip_seek_btree(cursor);
1498         KKASSERT(error == 0);
1499         cursor->key_beg = save;
1500
1501         /*
1502          * If the memory record was previous returned to
1503          * the caller and the media record matches
1504          * (-1/+1: only create_tid differs), then iterate
1505          * the media record to avoid a double result.
1506          */
1507         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1508             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1509                 elm = &cursor->node->ondisk->elms[cursor->index];
1510                 r = hammer_btree_cmp(&elm->base,
1511                                      &cursor->iprec->leaf.base);
1512                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1513                         if (r >= -1 && r <= 1) {
1514                                 kprintf("HAMMER: Debug: iterated after "
1515                                         "re-seek (asof r=%d)\n", r);
1516                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1517                                 again = 1;
1518                         }
1519                 } else {
1520                         if (r == 0) {
1521                                 kprintf("HAMMER: Debug: iterated after "
1522                                         "re-seek\n");
1523                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1524                                 again = 1;
1525                         }
1526                 }
1527         }
1528         return(again);
1529 }
1530
1531 /*
1532  * Locate the first record within the cursor's key_beg/key_end range,
1533  * restricted to a particular inode.  0 is returned on success, ENOENT
1534  * if no records matched the requested range, or some other error.
1535  *
1536  * When 0 is returned hammer_ip_next() may be used to iterate additional
1537  * records within the requested range.
1538  *
1539  * This function can return EDEADLK, requiring the caller to terminate
1540  * the cursor and try again.
1541  */
1542
1543 int
1544 hammer_ip_first(hammer_cursor_t cursor)
1545 {
1546         hammer_inode_t ip __debugvar = cursor->ip;
1547         int error;
1548
1549         KKASSERT(ip != NULL);
1550
1551         /*
1552          * Clean up fields and setup for merged scan
1553          */
1554         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1555
1556         /*
1557          * Search the in-memory record list (Red-Black tree).  Unlike the
1558          * B-Tree search, mem_first checks for records in the range.
1559          *
1560          * This function will setup both ATEMEM and MEMEOF properly for
1561          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1562          */
1563         hammer_mem_first(cursor);
1564
1565         /*
1566          * Detect generation changes during blockages, including
1567          * blockages which occur on the initial btree search.
1568          */
1569         cursor->rec_generation = cursor->ip->rec_generation;
1570
1571         /*
1572          * Initial search and result
1573          */
1574         error = _hammer_ip_seek_btree(cursor);
1575         if (error == 0)
1576                 error = hammer_ip_next(cursor);
1577
1578         return (error);
1579 }
1580
1581 /*
1582  * Retrieve the next record in a merged iteration within the bounds of the
1583  * cursor.  This call may be made multiple times after the cursor has been
1584  * initially searched with hammer_ip_first().
1585  *
1586  * There are numerous special cases in this code to deal with races between
1587  * in-memory records and on-media records.
1588  *
1589  * 0 is returned on success, ENOENT if no further records match the
1590  * requested range, or some other error code is returned.
1591  */
1592 int
1593 hammer_ip_next(hammer_cursor_t cursor)
1594 {
1595         hammer_btree_elm_t elm;
1596         hammer_record_t rec;
1597         hammer_record_t tmprec;
1598         int error;
1599         int r;
1600
1601 again:
1602         /*
1603          * Get the next on-disk record
1604          *
1605          * NOTE: If we deleted the last on-disk record we had scanned
1606          *       ATEDISK will be clear and RETEST will be set, forcing
1607          *       a call to iterate.  The fact that ATEDISK is clear causes
1608          *       iterate to re-test the 'current' element.  If ATEDISK is
1609          *       set, iterate will skip the 'current' element.
1610          */
1611         error = 0;
1612         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1613                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1614                                      HAMMER_CURSOR_RETEST)) {
1615                         error = hammer_btree_iterate(cursor);
1616                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1617                         if (error == 0) {
1618                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1619                                 hammer_cache_node(&cursor->ip->cache[1],
1620                                                   cursor->node);
1621                         } else if (error == ENOENT) {
1622                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1623                                                  HAMMER_CURSOR_ATEDISK;
1624                                 error = 0;
1625                         }
1626                 }
1627         }
1628
1629         /*
1630          * If the generation changed the backend has deleted or committed
1631          * one or more memory records since our last check.
1632          *
1633          * When this case occurs if the disk cursor is > current memory record
1634          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1635          * Since the cursor is ahead it must have not yet been eaten (if
1636          * not at eof anyway). (XXX data offset case?)
1637          *
1638          * NOTE: we are not doing a full check here.  That will be handled
1639          * later on.
1640          *
1641          * If we have exhausted all memory records we do not have to do any
1642          * further seeks.
1643          */
1644         while (cursor->rec_generation != cursor->ip->rec_generation &&
1645                error == 0
1646         ) {
1647                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1648                 cursor->rec_generation = cursor->ip->rec_generation;
1649                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1650                         break;
1651                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1652                         r = 1;
1653                 } else {
1654                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1655                         elm = &cursor->node->ondisk->elms[cursor->index];
1656                         r = hammer_btree_cmp(&elm->base,
1657                                              &cursor->iprec->leaf.base);
1658                 }
1659
1660                 /*
1661                  * Do we re-seek the media cursor?
1662                  */
1663                 if (r > 0) {
1664                         if (_hammer_ip_reseek(cursor))
1665                                 goto again;
1666                 }
1667         }
1668
1669         /*
1670          * We can now safely get the next in-memory record.  We cannot
1671          * block here.
1672          *
1673          * hammer_rec_scan_cmp:  Is the record still in our general range,
1674          *                       (non-inclusive of snapshot exclusions)?
1675          * hammer_rec_scan_callback: Is the record in our snapshot?
1676          */
1677         tmprec = NULL;
1678         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1679                 /*
1680                  * If the current memory record was eaten then get the next
1681                  * one.  Stale records are skipped.
1682                  */
1683                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1684                         tmprec = cursor->iprec;
1685                         cursor->iprec = NULL;
1686                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1687                         while (rec) {
1688                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1689                                         break;
1690                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1691                                         break;
1692                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1693                         }
1694                         if (cursor->iprec) {
1695                                 KKASSERT(cursor->iprec == rec);
1696                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1697                         } else {
1698                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1699                         }
1700                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1701                 }
1702         }
1703
1704         /*
1705          * MEMORY RECORD VALIDITY TEST
1706          *
1707          * (We still can't block, which is why tmprec is being held so
1708          * long).
1709          *
1710          * If the memory record is no longer valid we skip it.  It may
1711          * have been deleted by the frontend.  If it was deleted or
1712          * committed by the backend the generation change re-seeked the
1713          * disk cursor and the record will be present there.
1714          */
1715         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1716                 KKASSERT(cursor->iprec);
1717                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1718                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1719                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1720                         if (tmprec)
1721                                 hammer_rel_mem_record(tmprec);
1722                         goto again;
1723                 }
1724         }
1725         if (tmprec)
1726                 hammer_rel_mem_record(tmprec);
1727
1728         /*
1729          * Extract either the disk or memory record depending on their
1730          * relative position.
1731          */
1732         error = 0;
1733         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1734         case 0:
1735                 /*
1736                  * Both entries valid.   Compare the entries and nominally
1737                  * return the first one in the sort order.  Numerous cases
1738                  * require special attention, however.
1739                  */
1740                 elm = &cursor->node->ondisk->elms[cursor->index];
1741                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1742
1743                 /*
1744                  * If the two entries differ only by their key (-2/2) or
1745                  * create_tid (-1/1), and are DATA records, we may have a
1746                  * nominal match.  We have to calculate the base file
1747                  * offset of the data.
1748                  */
1749                 if (r <= 2 && r >= -2 && r != 0 &&
1750                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1751                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1752                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1753                         int64_t base2 = cursor->iprec->leaf.base.key -
1754                                         cursor->iprec->leaf.data_len;
1755                         if (base1 == base2)
1756                                 r = 0;
1757                 }
1758
1759                 if (r < 0) {
1760                         error = hammer_btree_extract(cursor,
1761                                                      HAMMER_CURSOR_GET_LEAF);
1762                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1763                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1764                         break;
1765                 }
1766
1767                 /*
1768                  * If the entries match exactly the memory entry is either
1769                  * an on-disk directory entry deletion or a bulk data
1770                  * overwrite.  If it is a directory entry deletion we eat
1771                  * both entries.
1772                  *
1773                  * For the bulk-data overwrite case it is possible to have
1774                  * visibility into both, which simply means the syncer
1775                  * hasn't gotten around to doing the delete+insert sequence
1776                  * on the B-Tree.  Use the memory entry and throw away the
1777                  * on-disk entry.
1778                  *
1779                  * If the in-memory record is not either of these we
1780                  * probably caught the syncer while it was syncing it to
1781                  * the media.  Since we hold a shared lock on the cursor,
1782                  * the in-memory record had better be marked deleted at
1783                  * this point.
1784                  */
1785                 if (r == 0) {
1786                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1787                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1788                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1789                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1790                                         goto again;
1791                                 }
1792                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1793                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1794                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1795                                 }
1796                                 /* fall through to memory entry */
1797                         } else {
1798                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1799                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1800                                 goto again;
1801                         }
1802                 }
1803                 /* fall through to the memory entry */
1804         case HAMMER_CURSOR_ATEDISK:
1805                 /*
1806                  * Only the memory entry is valid.
1807                  */
1808                 cursor->leaf = &cursor->iprec->leaf;
1809                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1810                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1811
1812                 /*
1813                  * If the memory entry is an on-disk deletion we should have
1814                  * also had found a B-Tree record.  If the backend beat us
1815                  * to it it would have interlocked the cursor and we should
1816                  * have seen the in-memory record marked DELETED_FE.
1817                  */
1818                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1819                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1820                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1821                 }
1822                 break;
1823         case HAMMER_CURSOR_ATEMEM:
1824                 /*
1825                  * Only the disk entry is valid
1826                  */
1827                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1828                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1829                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1830                 break;
1831         default:
1832                 /*
1833                  * Neither entry is valid
1834                  *
1835                  * XXX error not set properly
1836                  */
1837                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1838                 cursor->leaf = NULL;
1839                 error = ENOENT;
1840                 break;
1841         }
1842         return(error);
1843 }
1844
1845 /*
1846  * Resolve the cursor->data pointer for the current cursor position in
1847  * a merged iteration.
1848  */
1849 int
1850 hammer_ip_resolve_data(hammer_cursor_t cursor)
1851 {
1852         hammer_record_t record;
1853         int error;
1854
1855         if (hammer_cursor_inmem(cursor)) {
1856                 /*
1857                  * The data associated with an in-memory record is usually
1858                  * kmalloced, but reserve-ahead data records will have an
1859                  * on-disk reference.
1860                  *
1861                  * NOTE: Reserve-ahead data records must be handled in the
1862                  * context of the related high level buffer cache buffer
1863                  * to interlock against async writes.
1864                  */
1865                 record = cursor->iprec;
1866                 cursor->data = record->data;
1867                 error = 0;
1868                 if (cursor->data == NULL) {
1869                         KKASSERT(record->leaf.base.rec_type ==
1870                                  HAMMER_RECTYPE_DATA);
1871                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1872                                                     record->leaf.data_offset,
1873                                                     record->leaf.data_len,
1874                                                     &error,
1875                                                     &cursor->data_buffer);
1876                 }
1877         } else {
1878                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1879                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1880         }
1881         return(error);
1882 }
1883
1884 /*
1885  * Backend truncation / record replacement - delete records in range.
1886  *
1887  * Delete all records within the specified range for inode ip.  In-memory
1888  * records still associated with the frontend are ignored.
1889  *
1890  * If truncating is non-zero in-memory records associated with the back-end
1891  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1892  *
1893  * NOTES:
1894  *
1895  *      * An unaligned range will cause new records to be added to cover
1896  *        the edge cases. (XXX not implemented yet).
1897  *
1898  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1899  *        also do not deal with unaligned ranges.
1900  *
1901  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1902  *
1903  *      * Record keys for regular file data have to be special-cased since
1904  *        they indicate the end of the range (key = base + bytes).
1905  *
1906  *      * This function may be asked to delete ridiculously huge ranges, for
1907  *        example if someone truncates or removes a 1TB regular file.  We
1908  *        must be very careful on restarts and we may have to stop w/
1909  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1910  */
1911 int
1912 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1913                        int64_t ran_beg, int64_t ran_end, int truncating)
1914 {
1915         hammer_transaction_t trans = cursor->trans;
1916         hammer_btree_leaf_elm_t leaf;
1917         int error;
1918         int64_t off;
1919         int64_t tmp64;
1920
1921 #if 0
1922         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1923 #endif
1924
1925         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1926 retry:
1927         hammer_normalize_cursor(cursor);
1928         cursor->key_beg.localization = ip->obj_localization +
1929                                        HAMMER_LOCALIZE_MISC;
1930         cursor->key_beg.obj_id = ip->obj_id;
1931         cursor->key_beg.create_tid = 0;
1932         cursor->key_beg.delete_tid = 0;
1933         cursor->key_beg.obj_type = 0;
1934
1935         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1936                 cursor->key_beg.key = ran_beg;
1937                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1938         } else {
1939                 /*
1940                  * The key in the B-Tree is (base+bytes), so the first possible
1941                  * matching key is ran_beg + 1.
1942                  */
1943                 cursor->key_beg.key = ran_beg + 1;
1944                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1945         }
1946
1947         cursor->key_end = cursor->key_beg;
1948         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1949                 cursor->key_end.key = ran_end;
1950         } else {
1951                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1952                 if (tmp64 < ran_end)
1953                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1954                 else
1955                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1956         }
1957
1958         cursor->asof = ip->obj_asof;
1959         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1960         cursor->flags |= HAMMER_CURSOR_ASOF;
1961         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1962         cursor->flags |= HAMMER_CURSOR_BACKEND;
1963         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1964
1965         error = hammer_ip_first(cursor);
1966
1967         /*
1968          * Iterate through matching records and mark them as deleted.
1969          */
1970         while (error == 0) {
1971                 leaf = cursor->leaf;
1972
1973                 KKASSERT(leaf->base.delete_tid == 0);
1974                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1975
1976                 /*
1977                  * There may be overlap cases for regular file data.  Also
1978                  * remember the key for a regular file record is (base + len),
1979                  * NOT (base).
1980                  *
1981                  * Note that due to duplicates (mem & media) allowed by
1982                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1983                  */
1984                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1985                         off = leaf->base.key - leaf->data_len;
1986                         /*
1987                          * Check the left edge case.  We currently do not
1988                          * split existing records.
1989                          */
1990                         if (off < ran_beg && leaf->base.key > ran_beg) {
1991                                 panic("hammer left edge case %016llx %d\n",
1992                                         (long long)leaf->base.key,
1993                                         leaf->data_len);
1994                         }
1995
1996                         /*
1997                          * Check the right edge case.  Note that the
1998                          * record can be completely out of bounds, which
1999                          * terminates the search.
2000                          *
2001                          * base->key is exclusive of the right edge while
2002                          * ran_end is inclusive of the right edge.  The
2003                          * (key - data_len) left boundary is inclusive.
2004                          *
2005                          * XXX theory-check this test at some point, are
2006                          * we missing a + 1 somewhere?  Note that ran_end
2007                          * could overflow.
2008                          */
2009                         if (leaf->base.key - 1 > ran_end) {
2010                                 if (leaf->base.key - leaf->data_len > ran_end)
2011                                         break;
2012                                 panic("hammer right edge case\n");
2013                         }
2014                 } else {
2015                         off = leaf->base.key;
2016                 }
2017
2018                 /*
2019                  * Delete the record.  When truncating we do not delete
2020                  * in-memory (data) records because they represent data
2021                  * written after the truncation.
2022                  *
2023                  * This will also physically destroy the B-Tree entry and
2024                  * data if the retention policy dictates.  The function
2025                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2026                  * to retest the new 'current' element.
2027                  */
2028                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2029                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2030                         /*
2031                          * If we have built up too many meta-buffers we risk
2032                          * deadlocking the kernel and must stop.  This can
2033                          * occur when deleting ridiculously huge files.
2034                          * sync_trunc_off is updated so the next cycle does
2035                          * not re-iterate records we have already deleted.
2036                          *
2037                          * This is only done with formal truncations.
2038                          */
2039                         if (truncating > 1 && error == 0 &&
2040                             hammer_flusher_meta_limit(ip->hmp)) {
2041                                 ip->sync_trunc_off = off;
2042                                 error = EWOULDBLOCK;
2043                         }
2044                 }
2045                 if (error)
2046                         break;
2047                 ran_beg = off;  /* for restart */
2048                 error = hammer_ip_next(cursor);
2049         }
2050         if (cursor->node)
2051                 hammer_cache_node(&ip->cache[1], cursor->node);
2052
2053         if (error == EDEADLK) {
2054                 hammer_done_cursor(cursor);
2055                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2056                 if (error == 0)
2057                         goto retry;
2058         }
2059         if (error == ENOENT)
2060                 error = 0;
2061         return(error);
2062 }
2063
2064 /*
2065  * This backend function deletes the specified record on-disk, similar to
2066  * delete_range but for a specific record.  Unlike the exact deletions
2067  * used when deleting a directory entry this function uses an ASOF search
2068  * like delete_range.
2069  *
2070  * This function may be called with ip->obj_asof set for a slave snapshot,
2071  * so don't use it.  We always delete non-historical records only.
2072  */
2073 static int
2074 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2075                       hammer_btree_leaf_elm_t leaf)
2076 {
2077         hammer_transaction_t trans = cursor->trans;
2078         int error;
2079
2080         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2081 retry:
2082         hammer_normalize_cursor(cursor);
2083         cursor->key_beg = leaf->base;
2084         cursor->asof = HAMMER_MAX_TID;
2085         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2086         cursor->flags |= HAMMER_CURSOR_ASOF;
2087         cursor->flags |= HAMMER_CURSOR_BACKEND;
2088         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2089
2090         error = hammer_btree_lookup(cursor);
2091         if (error == 0) {
2092                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2093         }
2094         if (error == EDEADLK) {
2095                 hammer_done_cursor(cursor);
2096                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2097                 if (error == 0)
2098                         goto retry;
2099         }
2100         return(error);
2101 }
2102
2103 /*
2104  * This function deletes remaining auxillary records when an inode is
2105  * being deleted.  This function explicitly does not delete the
2106  * inode record, directory entry, data, or db records.  Those must be
2107  * properly disposed of prior to this call.
2108  */
2109 int
2110 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2111 {
2112         hammer_transaction_t trans = cursor->trans;
2113         hammer_btree_leaf_elm_t leaf;
2114         int error;
2115
2116         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2117 retry:
2118         hammer_normalize_cursor(cursor);
2119         cursor->key_beg.localization = ip->obj_localization +
2120                                        HAMMER_LOCALIZE_MISC;
2121         cursor->key_beg.obj_id = ip->obj_id;
2122         cursor->key_beg.create_tid = 0;
2123         cursor->key_beg.delete_tid = 0;
2124         cursor->key_beg.obj_type = 0;
2125         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2126         cursor->key_beg.key = HAMMER_MIN_KEY;
2127
2128         cursor->key_end = cursor->key_beg;
2129         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2130         cursor->key_end.key = HAMMER_MAX_KEY;
2131
2132         cursor->asof = ip->obj_asof;
2133         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2134         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2135         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2136         cursor->flags |= HAMMER_CURSOR_BACKEND;
2137
2138         error = hammer_ip_first(cursor);
2139
2140         /*
2141          * Iterate through matching records and mark them as deleted.
2142          */
2143         while (error == 0) {
2144                 leaf = cursor->leaf;
2145
2146                 KKASSERT(leaf->base.delete_tid == 0);
2147
2148                 /*
2149                  * Mark the record and B-Tree entry as deleted.  This will
2150                  * also physically delete the B-Tree entry, record, and
2151                  * data if the retention policy dictates.  The function
2152                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2153                  * to retest the new 'current' element.
2154                  *
2155                  * Directory entries (and delete-on-disk directory entries)
2156                  * must be synced and cannot be deleted.
2157                  */
2158                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2159                 ++*countp;
2160                 if (error)
2161                         break;
2162                 error = hammer_ip_next(cursor);
2163         }
2164         if (cursor->node)
2165                 hammer_cache_node(&ip->cache[1], cursor->node);
2166         if (error == EDEADLK) {
2167                 hammer_done_cursor(cursor);
2168                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2169                 if (error == 0)
2170                         goto retry;
2171         }
2172         if (error == ENOENT)
2173                 error = 0;
2174         return(error);
2175 }
2176
2177 /*
2178  * Delete the record at the current cursor.  On success the cursor will
2179  * be positioned appropriately for an iteration but may no longer be at
2180  * a leaf node.
2181  *
2182  * This routine is only called from the backend.
2183  *
2184  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2185  * cursor and retry.
2186  */
2187 int
2188 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2189                         hammer_tid_t tid)
2190 {
2191         hammer_record_t iprec;
2192         hammer_mount_t hmp;
2193         int error;
2194
2195         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2196         KKASSERT(tid != 0);
2197         hmp = cursor->node->hmp;
2198
2199         /*
2200          * In-memory (unsynchronized) records can simply be freed.  This
2201          * only occurs in range iterations since all other records are
2202          * individually synchronized.  Thus there should be no confusion with
2203          * the interlock.
2204          *
2205          * An in-memory record may be deleted before being committed to disk,
2206          * but could have been accessed in the mean time.  The reservation
2207          * code will deal with the case.
2208          */
2209         if (hammer_cursor_inmem(cursor)) {
2210                 iprec = cursor->iprec;
2211                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2212                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2213                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2214                 KKASSERT(iprec->ip == ip);
2215                 ++ip->rec_generation;
2216                 return(0);
2217         }
2218
2219         /*
2220          * On-disk records are marked as deleted by updating their delete_tid.
2221          * This does not effect their position in the B-Tree (which is based
2222          * on their create_tid).
2223          *
2224          * Frontend B-Tree operations track inodes so we tell
2225          * hammer_delete_at_cursor() not to.
2226          */
2227         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2228
2229         if (error == 0) {
2230                 error = hammer_delete_at_cursor(
2231                                 cursor,
2232                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2233                                 cursor->trans->tid,
2234                                 cursor->trans->time32,
2235                                 0, NULL);
2236         }
2237         return(error);
2238 }
2239
2240 /*
2241  * Used to write a generic record w/optional data to the media b-tree
2242  * when no inode context is available.  Used by the mirroring and
2243  * snapshot code.
2244  *
2245  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2246  * flagged for backend operation and not flagged ASOF (since we are
2247  * doing an insertion).
2248  *
2249  * This function will acquire the appropriate sync lock and will set
2250  * the cursor insertion flag for the operation, do the btree lookup,
2251  * and the insertion, and clear the insertion flag and sync lock before
2252  * returning.  The cursor state will be such that the caller can continue
2253  * scanning (used by the mirroring code).
2254  *
2255  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2256  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2257  *
2258  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2259  *                retry.
2260  *
2261  *       EALREADY can be returned if the record already exists (WARNING,
2262  *                because ASOF cannot be used no check is made for illegal
2263  *                duplicates).
2264  *
2265  * NOTE: Do not use the function for normal inode-related records as this
2266  *       functions goes directly to the media and is not integrated with
2267  *       in-memory records.
2268  */
2269 int
2270 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2271                         void *udata, int mode)
2272 {
2273         hammer_transaction_t trans;
2274         hammer_buffer_t data_buffer;
2275         hammer_off_t ndata_offset;
2276         hammer_tid_t high_tid;
2277         void *ndata;
2278         int error;
2279         int doprop;
2280
2281         trans = cursor->trans;
2282         data_buffer = NULL;
2283         ndata_offset = 0;
2284         doprop = 0;
2285
2286         KKASSERT((cursor->flags &
2287                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2288                   (HAMMER_CURSOR_BACKEND));
2289
2290         hammer_sync_lock_sh(trans);
2291
2292         if (leaf->data_len) {
2293                 ndata = hammer_alloc_data(trans, leaf->data_len,
2294                                           leaf->base.rec_type,
2295                                           &ndata_offset, &data_buffer,
2296                                           0, &error);
2297                 if (ndata == NULL) {
2298                         hammer_sync_unlock(trans);
2299                         return (error);
2300                 }
2301                 leaf->data_offset = ndata_offset;
2302                 hammer_modify_buffer(trans, data_buffer, NULL, 0);
2303
2304                 switch(mode) {
2305                 case HAMMER_CREATE_MODE_UMIRROR:
2306                         error = copyin(udata, ndata, leaf->data_len);
2307                         if (error == 0) {
2308                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2309                                         kprintf("data crc mismatch on pipe\n");
2310                                         error = EINVAL;
2311                                 } else {
2312                                         error = hammer_cursor_localize_data(
2313                                                         ndata, leaf);
2314                                 }
2315                         }
2316                         break;
2317                 case HAMMER_CREATE_MODE_SYS:
2318                         bcopy(udata, ndata, leaf->data_len);
2319                         error = 0;
2320                         hammer_crc_set_leaf(ndata, leaf);
2321                         break;
2322                 default:
2323                         panic("hammer: hammer_create_at_cursor: bad mode %d",
2324                                 mode);
2325                         break; /* NOT REACHED */
2326                 }
2327                 hammer_modify_buffer_done(data_buffer);
2328         } else {
2329                 leaf->data_offset = 0;
2330                 error = 0;
2331                 ndata = NULL;
2332         }
2333         if (error)
2334                 goto failed;
2335
2336         /*
2337          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2338          */
2339         cursor->flags |= HAMMER_CURSOR_INSERT;
2340         error = hammer_btree_lookup(cursor);
2341         if (error != ENOENT) {
2342                 if (error == 0)
2343                         error = EALREADY;
2344                 goto failed;
2345         }
2346         error = hammer_btree_insert(cursor, leaf, &doprop);
2347
2348         /*
2349          * Cursor is left on current element, we want to skip it now.
2350          * (in case the caller is scanning)
2351          */
2352         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2353         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2354
2355         /*
2356          * If the insertion happens to be creating (and not just replacing)
2357          * an inode we have to track it.
2358          */
2359         if (error == 0 &&
2360             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2361             leaf->base.delete_tid == 0) {
2362                 hammer_modify_volume_field(trans, trans->rootvol,
2363                                            vol0_stat_inodes);
2364                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2365                 hammer_modify_volume_done(trans->rootvol);
2366         }
2367
2368         /*
2369          * vol0_next_tid must track the highest TID stored in the filesystem.
2370          * We do not need to generate undo for this update.
2371          */
2372         high_tid = leaf->base.create_tid;
2373         if (high_tid < leaf->base.delete_tid)
2374                 high_tid = leaf->base.delete_tid;
2375         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2376                 hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2377                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2378                 hammer_modify_volume_done(trans->rootvol);
2379         }
2380
2381         /*
2382          * WARNING!  cursor's leaf pointer may have changed after
2383          *           do_propagation returns.
2384          */
2385         if (error == 0 && doprop)
2386                 hammer_btree_do_propagation(cursor, NULL, leaf);
2387
2388 failed:
2389         /*
2390          * Cleanup
2391          */
2392         if (error && leaf->data_offset) {
2393                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2394
2395         }
2396         hammer_sync_unlock(trans);
2397         if (data_buffer)
2398                 hammer_rel_buffer(data_buffer, 0);
2399         return (error);
2400 }
2401
2402 /*
2403  * Delete the B-Tree element at the current cursor and do any necessary
2404  * mirror propagation.
2405  *
2406  * The cursor must be properly positioned for an iteration on return but
2407  * may be pointing at an internal element.
2408  *
2409  * An element can be un-deleted by passing a delete_tid of 0 with
2410  * HAMMER_DELETE_ADJUST.
2411  */
2412 int
2413 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2414                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2415                         int track, int64_t *stat_bytes)
2416 {
2417         struct hammer_btree_leaf_elm save_leaf;
2418         hammer_transaction_t trans;
2419         hammer_btree_leaf_elm_t leaf;
2420         hammer_node_t node;
2421         hammer_btree_elm_t elm;
2422         hammer_off_t data_offset;
2423         int32_t data_len;
2424         u_int16_t rec_type;
2425         int error;
2426         int icount;
2427         int doprop;
2428
2429         error = hammer_cursor_upgrade(cursor);
2430         if (error)
2431                 return(error);
2432
2433         trans = cursor->trans;
2434         node = cursor->node;
2435         elm = &node->ondisk->elms[cursor->index];
2436         leaf = &elm->leaf;
2437         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2438
2439         hammer_sync_lock_sh(trans);
2440         doprop = 0;
2441         icount = 0;
2442
2443         /*
2444          * Adjust the delete_tid.  Update the mirror_tid propagation field
2445          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2446          */
2447         if (delete_flags & HAMMER_DELETE_ADJUST) {
2448                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2449                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2450                                 icount = -1;
2451                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2452                                 icount = 1;
2453                 }
2454
2455                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2456                 elm->leaf.base.delete_tid = delete_tid;
2457                 elm->leaf.delete_ts = delete_ts;
2458                 hammer_modify_node_done(node);
2459
2460                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2461                         hammer_modify_node_field(trans, node, mirror_tid);
2462                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2463                         hammer_modify_node_done(node);
2464                         doprop = 1;
2465                         if (hammer_debug_general & 0x0002) {
2466                                 kprintf("delete_at_cursor: propagate %016llx"
2467                                         " @%016llx\n",
2468                                         (long long)elm->leaf.base.delete_tid,
2469                                         (long long)node->node_offset);
2470                         }
2471                 }
2472
2473                 /*
2474                  * Adjust for the iteration.  We have deleted the current
2475                  * element and want to clear ATEDISK so the iteration does
2476                  * not skip the element after, which now becomes the current
2477                  * element.  This element must be re-tested if doing an
2478                  * iteration, which is handled by the RETEST flag.
2479                  */
2480                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2481                         cursor->flags |= HAMMER_CURSOR_RETEST;
2482                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2483                 }
2484
2485                 /*
2486                  * An on-disk record cannot have the same delete_tid
2487                  * as its create_tid.  In a chain of record updates
2488                  * this could result in a duplicate record.
2489                  */
2490                 KKASSERT(elm->leaf.base.delete_tid !=
2491                          elm->leaf.base.create_tid);
2492         }
2493
2494         /*
2495          * Destroy the B-Tree element if asked (typically if a nohistory
2496          * file or mount, or when called by the pruning code).
2497          *
2498          * Adjust the ATEDISK flag to properly support iterations.
2499          */
2500         if (delete_flags & HAMMER_DELETE_DESTROY) {
2501                 data_offset = elm->leaf.data_offset;
2502                 data_len = elm->leaf.data_len;
2503                 rec_type = elm->leaf.base.rec_type;
2504                 if (doprop) {
2505                         save_leaf = elm->leaf;
2506                         leaf = &save_leaf;
2507                 }
2508                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2509                     elm->leaf.base.delete_tid == 0) {
2510                         icount = -1;
2511                 }
2512
2513                 error = hammer_btree_delete(cursor);
2514                 if (error == 0) {
2515                         /*
2516                          * The deletion moves the next element (if any) to
2517                          * the current element position.  We must clear
2518                          * ATEDISK so this element is not skipped and we
2519                          * must set RETEST to force any iteration to re-test
2520                          * the element.
2521                          */
2522                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2523                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2524                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2525                         }
2526                 }
2527                 if (error == 0) {
2528                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2529                         case HAMMER_ZONE_LARGE_DATA:
2530                         case HAMMER_ZONE_SMALL_DATA:
2531                         case HAMMER_ZONE_META:
2532                                 hammer_blockmap_free(trans,
2533                                                      data_offset, data_len);
2534                                 break;
2535                         default:
2536                                 break;
2537                         }
2538                 }
2539         }
2540
2541         /*
2542          * Track inode count and next_tid.  This is used by the mirroring
2543          * and PFS code.  icount can be negative, zero, or positive.
2544          */
2545         if (error == 0 && track) {
2546                 if (icount) {
2547                         hammer_modify_volume_field(trans, trans->rootvol,
2548                                                    vol0_stat_inodes);
2549                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2550                         hammer_modify_volume_done(trans->rootvol);
2551                 }
2552                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2553                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2554                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2555                         hammer_modify_volume_done(trans->rootvol);
2556                 }
2557         }
2558
2559         /*
2560          * mirror_tid propagation occurs if the node's mirror_tid had to be
2561          * updated while adjusting the delete_tid.
2562          *
2563          * This occurs when deleting even in nohistory mode, but does not
2564          * occur when pruning an already-deleted node.
2565          *
2566          * cursor->ip is NULL when called from the pruning, mirroring,
2567          * and pfs code.  If non-NULL propagation will be conditionalized
2568          * on whether the PFS is in no-history mode or not.
2569          *
2570          * WARNING: cursor's leaf pointer may have changed after do_propagation
2571          *          returns!
2572          */
2573         if (doprop) {
2574                 if (cursor->ip)
2575                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2576                 else
2577                         hammer_btree_do_propagation(cursor, NULL, leaf);
2578         }
2579         hammer_sync_unlock(trans);
2580         return (error);
2581 }
2582
2583 /*
2584  * Determine whether we can remove a directory.  This routine checks whether
2585  * a directory is empty or not and enforces flush connectivity.
2586  *
2587  * Flush connectivity requires that we block if the target directory is
2588  * currently flushing, otherwise it may not end up in the same flush group.
2589  *
2590  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2591  */
2592 int
2593 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2594 {
2595         struct hammer_cursor cursor;
2596         int error;
2597
2598         /*
2599          * Check directory empty
2600          */
2601         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2602
2603         cursor.key_beg.localization = ip->obj_localization +
2604                                       hammer_dir_localization(ip);
2605         cursor.key_beg.obj_id = ip->obj_id;
2606         cursor.key_beg.create_tid = 0;
2607         cursor.key_beg.delete_tid = 0;
2608         cursor.key_beg.obj_type = 0;
2609         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2610         cursor.key_beg.key = HAMMER_MIN_KEY;
2611
2612         cursor.key_end = cursor.key_beg;
2613         cursor.key_end.rec_type = 0xFFFF;
2614         cursor.key_end.key = HAMMER_MAX_KEY;
2615
2616         cursor.asof = ip->obj_asof;
2617         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2618
2619         error = hammer_ip_first(&cursor);
2620         if (error == ENOENT)
2621                 error = 0;
2622         else if (error == 0)
2623                 error = ENOTEMPTY;
2624         hammer_done_cursor(&cursor);
2625         return(error);
2626 }
2627
2628 /*
2629  * Localize the data payload.  Directory entries may need their
2630  * localization adjusted.
2631  */
2632 static
2633 int
2634 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2635                             hammer_btree_leaf_elm_t leaf)
2636 {
2637         u_int32_t localization;
2638
2639         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2640                 localization = leaf->base.localization &
2641                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2642                 if (data->entry.localization != localization) {
2643                         data->entry.localization = localization;
2644                         hammer_crc_set_leaf(data, leaf);
2645                 }
2646         }
2647         return(0);
2648 }