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
  35 #include "hammer.h"
  36
  37 static int hammer_mem_lookup(hammer_cursor_t cursor);
  38 static int hammer_mem_first(hammer_cursor_t cursor);
  39 static int hammer_frontend_trunc_callback(hammer_record_t record,
  40                                 void *data __unused);
  41 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
  42 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  43 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
  44                                 hammer_btree_leaf_elm_t leaf);
  45 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
  46                                 hammer_btree_leaf_elm_t leaf);
  47
  48 struct rec_trunc_info {
  49         uint16_t        rec_type;
  50         int64_t         trunc_off;
  51 };
  52
  53 struct hammer_bulk_info {
  54         hammer_record_t record;
  55         hammer_record_t conflict;
  56 };
  57
  58 /*
  59  * Red-black tree support.  Comparison code for insertion.
  60  */
  61 static int
  62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  63 {
  64         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  65                 return(-1);
  66         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  67                 return(1);
  68
  69         if (rec1->leaf.base.key < rec2->leaf.base.key)
  70                 return(-1);
  71         if (rec1->leaf.base.key > rec2->leaf.base.key)
  72                 return(1);
  73
  74         /*
  75          * For search & insertion purposes records deleted by the
  76          * frontend or deleted/committed by the backend are silently
  77          * ignored.  Otherwise pipelined insertions will get messed
  78          * up.
  79          *
  80          * rec1 is greater then rec2 if rec1 is marked deleted.
  81          * rec1 is less then rec2 if rec2 is marked deleted.
  82          *
  83          * Multiple deleted records may be present, do not return 0
  84          * if both are marked deleted.
  85          */
  86         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  87                            HAMMER_RECF_COMMITTED)) {
  88                 return(1);
  89         }
  90         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  91                            HAMMER_RECF_COMMITTED)) {
  92                 return(-1);
  93         }
  94
  95         return(0);
  96 }
  97
  98 /*
  99  * Basic record comparison code similar to hammer_btree_cmp().
 100  *
 101  * obj_id is not compared and may not yet be assigned in the record.
 102  */
 103 static int
 104 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 105 {
 106         if (elm->rec_type < rec->leaf.base.rec_type)
 107                 return(-3);
 108         if (elm->rec_type > rec->leaf.base.rec_type)
 109                 return(3);
 110
 111         if (elm->key < rec->leaf.base.key)
 112                 return(-2);
 113         if (elm->key > rec->leaf.base.key)
 114                 return(2);
 115
 116         /*
 117          * Never match against an item deleted by the frontend
 118          * or backend, or committed by the backend.
 119          *
 120          * elm is less then rec if rec is marked deleted.
 121          */
 122         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 123                           HAMMER_RECF_COMMITTED)) {
 124                 return(-1);
 125         }
 126         return(0);
 127 }
 128
 129 /*
 130  * Ranged scan to locate overlapping record(s).  This is used by
 131  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 132  * to use a ranged scan because the keys for data records with the
 133  * same file base offset can be different due to differing data_len's.
 134  *
 135  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 136  */
 137 static int
 138 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 139 {
 140         struct hammer_bulk_info *info = data;
 141         hammer_btree_leaf_elm_t leaf = &info->record->leaf;
 142
 143         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 144                 return(-3);
 145         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 146                 return(3);
 147
 148         /*
 149          * Overlap compare
 150          */
 151         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 152                 /* rec_beg >= leaf_end */
 153                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 154                         return(2);
 155                 /* rec_end <= leaf_beg */
 156                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 157                         return(-2);
 158         } else {
 159                 if (rec->leaf.base.key < leaf->base.key)
 160                         return(-2);
 161                 if (rec->leaf.base.key > leaf->base.key)
 162                         return(2);
 163         }
 164
 165         /*
 166          * We have to return 0 at this point, even if DELETED_FE is set,
 167          * because returning anything else will cause the scan to ignore
 168          * one of the branches when we really want it to check both.
 169          */
 170         return(0);
 171 }
 172
 173 /*
 174  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 175  * is reversed so the comparison result has to be negated.  key_beg and
 176  * key_end are both range-inclusive.
 177  *
 178  * Localized deletions are not cached in-memory.
 179  */
 180 static
 181 int
 182 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 183 {
 184         hammer_cursor_t cursor = data;
 185         int r;
 186
 187         r = hammer_rec_cmp(&cursor->key_beg, rec);
 188         if (r > 1)
 189                 return(-1);
 190         r = hammer_rec_cmp(&cursor->key_end, rec);
 191         if (r < -1)
 192                 return(1);
 193         return(0);
 194 }
 195
 196 /*
 197  * This compare function is used when simply looking up key_beg.
 198  */
 199 static
 200 int
 201 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 202 {
 203         hammer_cursor_t cursor = data;
 204         int r;
 205
 206         r = hammer_rec_cmp(&cursor->key_beg, rec);
 207         if (r > 1)
 208                 return(-1);
 209         if (r < -1)
 210                 return(1);
 211         return(0);
 212 }
 213
 214 /*
 215  * Locate blocks within the truncation range.  Partial blocks do not count.
 216  */
 217 static
 218 int
 219 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 220 {
 221         struct rec_trunc_info *info = data;
 222
 223         if (rec->leaf.base.rec_type < info->rec_type)
 224                 return(-1);
 225         if (rec->leaf.base.rec_type > info->rec_type)
 226                 return(1);
 227
 228         switch(rec->leaf.base.rec_type) {
 229         case HAMMER_RECTYPE_DB:
 230                 /*
 231                  * DB record key is not beyond the truncation point, retain.
 232                  */
 233                 if (rec->leaf.base.key < info->trunc_off)
 234                         return(-1);
 235                 break;
 236         case HAMMER_RECTYPE_DATA:
 237                 /*
 238                  * DATA record offset start is not beyond the truncation point,
 239                  * retain.
 240                  */
 241                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 242                         return(-1);
 243                 break;
 244         default:
 245                 hpanic("unexpected record type");
 246         }
 247
 248         /*
 249          * The record start is >= the truncation point, return match,
 250          * the record should be destroyed.
 251          */
 252         return(0);
 253 }
 254
 255 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 256
 257 /*
 258  * Allocate a record for the caller to finish filling in.  The record is
 259  * returned referenced.  In order to manually set data call this function
 260  * with data_len=0 and then manually set record->leaf.data_len and
 261  * record->data later.
 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(hammer_record_t 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 RB-Tree
 411                          */
 412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 413                                 RB_REMOVE(hammer_rec_rb_tree,
 414                                           &ip->rec_tree,
 415                                           record);
 416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 417                                 KKASSERT(ip->rsv_recs > 0);
 418                                 if (RB_EMPTY(&ip->rec_tree)) {
 419                                         ip->flags &= ~HAMMER_INODE_XDIRTY;
 420                                         ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 421                                 }
 422                                 diddrop = 1;
 423                         } else {
 424                                 diddrop = 0;
 425                         }
 426
 427                         /*
 428                          * We must wait for any direct-IO to complete before
 429                          * we can destroy the record because the bio may
 430                          * have a reference to it.
 431                          */
 432                         if (record->gflags &
 433                            (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
 434                                 hammer_io_direct_wait(record);
 435                         }
 436
 437                         /*
 438                          * Account for the completion after the direct IO
 439                          * has completed.
 440                          */
 441                         if (diddrop) {
 442                                 --hmp->rsv_recs;
 443                                 --ip->rsv_recs;
 444                                 hmp->rsv_databytes -= record->leaf.data_len;
 445
 446                                 if (RB_EMPTY(&ip->rec_tree))
 447                                         hammer_test_inode(ip);
 448                                 if ((ip->flags & HAMMER_INODE_RECSW) &&
 449                                     ip->rsv_recs <= hammer_limit_inode_recs/2) {
 450                                         ip->flags &= ~HAMMER_INODE_RECSW;
 451                                         wakeup(&ip->rsv_recs);
 452                                 }
 453                         }
 454
 455                         /*
 456                          * Do this test after removing record from the RB-Tree.
 457                          */
 458                         if (target_ip) {
 459                                 hammer_test_inode(target_ip);
 460                                 hammer_rel_inode(target_ip, 0);
 461                         }
 462
 463                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 464                                 --hammer_count_record_datas;
 465                                 kfree(record->data, hmp->m_misc);
 466                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 467                         }
 468
 469                         /*
 470                          * Release the reservation.
 471                          *
 472                          * If the record was not committed we can theoretically
 473                          * undo the reservation.  However, doing so might
 474                          * create weird edge cases with the ordering of
 475                          * direct writes because the related buffer cache
 476                          * elements are per-vnode.  So we don't try.
 477                          */
 478                         if ((resv = record->resv) != NULL) {
 479                                 /* XXX undo leaf.data_offset,leaf.data_len */
 480                                 hammer_blockmap_reserve_complete(hmp, resv);
 481                                 record->resv = NULL;
 482                         }
 483                         record->data = NULL;
 484                         --hammer_count_records;
 485                         kfree(record, hmp->m_misc);
 486                 }
 487         }
 488 }
 489
 490 /*
 491  * Record visibility depends on whether the record is being accessed by
 492  * the backend or the frontend.  Backend tests ignore the frontend delete
 493  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 494  * must also check for commit races.
 495  *
 496  * Return non-zero if the record is visible, zero if it isn't or if it is
 497  * deleted.  Returns 0 if the record has been comitted (unless the special
 498  * delete-visibility flag is set).  A committed record must be located
 499  * via the media B-Tree.  Returns non-zero if the record is good.
 500  *
 501  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 502  * records to be returned.  This is so pending deletions are detected
 503  * when using an iterator to locate an unused hash key, or when we need
 504  * to locate historical records on-disk to destroy.
 505  */
 506 static __inline
 507 int
 508 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 509 {
 510         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 511                 return(1);
 512         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 513                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 514                                      HAMMER_RECF_COMMITTED)) {
 515                         return(0);
 516                 }
 517         } else {
 518                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 519                                      HAMMER_RECF_DELETED_BE |
 520                                      HAMMER_RECF_COMMITTED)) {
 521                         return(0);
 522                 }
 523         }
 524         return(1);
 525 }
 526
 527 /*
 528  * This callback is used as part of the RB_SCAN function for in-memory
 529  * records.  We terminate it (return -1) as soon as we get a match.
 530  *
 531  * This routine is used by frontend code.
 532  *
 533  * The primary compare code does not account for ASOF lookups.  This
 534  * code handles that case as well as a few others.
 535  */
 536 static
 537 int
 538 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 539 {
 540         hammer_cursor_t cursor = data;
 541
 542         /*
 543          * We terminate on success, so this should be NULL on entry.
 544          */
 545         KKASSERT(cursor->iprec == NULL);
 546
 547         /*
 548          * Skip if the record was marked deleted or committed.
 549          */
 550         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 551                 return(0);
 552
 553         /*
 554          * Skip if not visible due to our as-of TID
 555          */
 556         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 557                 if (cursor->asof < rec->leaf.base.create_tid)
 558                         return(0);
 559                 if (rec->leaf.base.delete_tid &&
 560                     cursor->asof >= rec->leaf.base.delete_tid) {
 561                         return(0);
 562                 }
 563         }
 564
 565         /*
 566          * ref the record.  The record is protected from backend B-Tree
 567          * interactions by virtue of the cursor's IP lock.
 568          */
 569         hammer_ref(&rec->lock);
 570
 571         /*
 572          * The record may have been deleted or committed while we
 573          * were blocked.  XXX remove?
 574          */
 575         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 576                 hammer_rel_mem_record(rec);
 577                 return(0);
 578         }
 579
 580         /*
 581          * Set the matching record and stop the scan.
 582          */
 583         cursor->iprec = rec;
 584         return(-1);
 585 }
 586
 587
 588 /*
 589  * Lookup an in-memory record given the key specified in the cursor.  Works
 590  * just like hammer_btree_lookup() but operates on an inode's in-memory
 591  * record list.
 592  *
 593  * The lookup must fail if the record is marked for deferred deletion.
 594  *
 595  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 596  */
 597 static
 598 int
 599 hammer_mem_lookup(hammer_cursor_t cursor)
 600 {
 601         KKASSERT(cursor->ip != NULL);
 602         if (cursor->iprec) {
 603                 hammer_rel_mem_record(cursor->iprec);
 604                 cursor->iprec = NULL;
 605         }
 606         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 607                                    hammer_rec_scan_callback, cursor);
 608
 609         return (cursor->iprec ? 0 : ENOENT);
 610 }
 611
 612 /*
 613  * hammer_mem_first() - locate the first in-memory record matching the
 614  * cursor within the bounds of the key range.
 615  *
 616  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 617  * will set ATEMEM the same as MEMEOF, and does not return any error.
 618  */
 619 static
 620 int
 621 hammer_mem_first(hammer_cursor_t cursor)
 622 {
 623         KKASSERT(cursor->ip != NULL);
 624         if (cursor->iprec) {
 625                 hammer_rel_mem_record(cursor->iprec);
 626                 cursor->iprec = NULL;
 627         }
 628         hammer_rec_rb_tree_RB_SCAN(&cursor->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         return (cursor->iprec ? 0 : ENOENT);
 637 }
 638
 639 /************************************************************************
 640  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 641  ************************************************************************
 642  *
 643  * These functions manipulate in-memory records.  Such records typically
 644  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 645  */
 646
 647 /*
 648  * Add a directory entry (dip,ncp) which references inode (ip).
 649  *
 650  * Note that the low 32 bits of the namekey are set temporarily to create
 651  * a unique in-memory record, and may be modified a second time when the
 652  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 653  * all 0's when synching to disk, which is not handled here.
 654  *
 655  * NOTE: bytes does not include any terminating \0 on name, and name might
 656  * not be terminated.
 657  */
 658 int
 659 hammer_ip_add_direntry(hammer_transaction_t trans,
 660                      hammer_inode_t dip, const char *name, int bytes,
 661                      hammer_inode_t ip)
 662 {
 663         struct hammer_cursor cursor;
 664         hammer_record_t record;
 665         int error;
 666         uint32_t max_iterations;
 667
 668         KKASSERT(dip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY);
 669
 670         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 671
 672         record->type = HAMMER_MEM_RECORD_ADD;
 673         record->leaf.base.localization = dip->obj_localization |
 674                                          hammer_dir_localization(dip);
 675         record->leaf.base.obj_id = dip->obj_id;
 676         record->leaf.base.key = hammer_direntry_namekey(dip, name, bytes,
 677                                                          &max_iterations);
 678         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 679         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 680         record->data->entry.obj_id = ip->obj_id;
 681         record->data->entry.localization = ip->obj_localization;
 682         bcopy(name, record->data->entry.name, bytes);
 683
 684         ++ip->ino_data.nlinks;
 685         ip->ino_data.ctime = trans->time;
 686         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 687
 688         /*
 689          * Find an unused namekey.  Both the in-memory record tree and
 690          * the B-Tree are checked.  We do not want historically deleted
 691          * names to create a collision as our iteration space may be limited,
 692          * and since create_tid wouldn't match anyway an ASOF search
 693          * must be used to locate collisions.
 694          *
 695          * delete-visibility is set so pending deletions do not give us
 696          * a false-negative on our ability to use an iterator.
 697          *
 698          * The iterator must not rollover the key.  Directory keys only
 699          * use the positive key space.
 700          */
 701         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 702         cursor.key_beg = record->leaf.base;
 703         cursor.flags |= HAMMER_CURSOR_ASOF;
 704         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 705         cursor.asof = ip->obj_asof;
 706
 707         while (hammer_ip_lookup(&cursor) == 0) {
 708                 ++record->leaf.base.key;
 709                 KKASSERT(record->leaf.base.key > 0);
 710                 cursor.key_beg.key = record->leaf.base.key;
 711                 if (--max_iterations == 0) {
 712                         hammer_rel_mem_record(record);
 713                         hmkprintf(trans->hmp, "Failed to find an unused namekey\n");
 714                         error = ENOSPC;
 715                         goto failed;
 716                 }
 717         }
 718
 719         /*
 720          * The target inode and the directory entry are bound together.
 721          */
 722         record->target_ip = ip;
 723         record->flush_state = HAMMER_FST_SETUP;
 724         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 725
 726         /*
 727          * The inode now has a dependancy and must be taken out of the idle
 728          * state.  An inode not in an idle state is given an extra reference.
 729          *
 730          * When transitioning to a SETUP state flag for an automatic reflush
 731          * when the dependancies are disposed of if someone is waiting on
 732          * the inode.
 733          */
 734         if (ip->flush_state == HAMMER_FST_IDLE) {
 735                 hammer_ref(&ip->lock);
 736                 ip->flush_state = HAMMER_FST_SETUP;
 737                 if (ip->flags & HAMMER_INODE_FLUSHW)
 738                         ip->flags |= HAMMER_INODE_REFLUSH;
 739         }
 740         error = hammer_mem_add(record);
 741         if (error == 0) {
 742                 dip->ino_data.mtime = trans->time;
 743                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 744         }
 745 failed:
 746         hammer_done_cursor(&cursor);
 747         return(error);
 748 }
 749
 750 /*
 751  * Delete the directory entry and update the inode link count.  The
 752  * cursor must be seeked to the directory entry record being deleted.
 753  *
 754  * The related inode should be share-locked by the caller.  The caller is
 755  * on the frontend.  It could also be NULL indicating that the directory
 756  * entry being removed has no related inode.
 757  *
 758  * This function can return EDEADLK requiring the caller to terminate
 759  * the cursor, any locks, wait on the returned record, and retry.
 760  */
 761 int
 762 hammer_ip_del_direntry(hammer_transaction_t trans,
 763                      hammer_cursor_t cursor, hammer_inode_t dip,
 764                      hammer_inode_t ip)
 765 {
 766         hammer_record_t record;
 767         int error;
 768
 769         if (hammer_cursor_inmem(cursor)) {
 770                 /*
 771                  * In-memory (unsynchronized) records can simply be freed.
 772                  *
 773                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 774                  * by the backend, we must still avoid races against the
 775                  * backend potentially syncing the record to the media.
 776                  *
 777                  * We cannot call hammer_ip_delete_record(), that routine may
 778                  * only be called from the backend.
 779                  */
 780                 record = cursor->iprec;
 781                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 782                                      HAMMER_RECF_DELETED_BE |
 783                                      HAMMER_RECF_COMMITTED)) {
 784                         KKASSERT(cursor->deadlk_rec == NULL);
 785                         hammer_ref(&record->lock);
 786                         cursor->deadlk_rec = record;
 787                         error = EDEADLK;
 788                 } else {
 789                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 790                         record->flags |= HAMMER_RECF_DELETED_FE;
 791                         error = 0;
 792                 }
 793         } else {
 794                 /*
 795                  * If the record is on-disk we have to queue the deletion by
 796                  * the record's key.  This also causes lookups to skip the
 797                  * record (lookups for the purposes of finding an unused
 798                  * directory key do not skip the record).
 799                  */
 800                 KKASSERT(dip->flags &
 801                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 802                 record = hammer_alloc_mem_record(dip, 0);
 803                 record->type = HAMMER_MEM_RECORD_DEL;
 804                 record->leaf.base = cursor->leaf->base;
 805                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
 806
 807                 /*
 808                  * ip may be NULL, indicating the deletion of a directory
 809                  * entry which has no related inode.
 810                  */
 811                 record->target_ip = ip;
 812                 if (ip) {
 813                         record->flush_state = HAMMER_FST_SETUP;
 814                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 815                                           target_entry);
 816                 } else {
 817                         record->flush_state = HAMMER_FST_IDLE;
 818                 }
 819
 820                 /*
 821                  * The inode now has a dependancy and must be taken out of
 822                  * the idle state.  An inode not in an idle state is given
 823                  * an extra reference.
 824                  *
 825                  * When transitioning to a SETUP state flag for an automatic
 826                  * reflush when the dependancies are disposed of if someone
 827                  * is waiting on the inode.
 828                  */
 829                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 830                         hammer_ref(&ip->lock);
 831                         ip->flush_state = HAMMER_FST_SETUP;
 832                         if (ip->flags & HAMMER_INODE_FLUSHW)
 833                                 ip->flags |= HAMMER_INODE_REFLUSH;
 834                 }
 835
 836                 error = hammer_mem_add(record);
 837         }
 838
 839         /*
 840          * One less link.  The file may still be open in the OS even after
 841          * all links have gone away.
 842          *
 843          * We have to terminate the cursor before syncing the inode to
 844          * avoid deadlocking against ourselves.  XXX this may no longer
 845          * be true.
 846          *
 847          * If nlinks drops to zero and the vnode is inactive (or there is
 848          * no vnode), call hammer_inode_unloadable_check() to zonk the
 849          * inode.  If we don't do this here the inode will not be destroyed
 850          * on-media until we unmount.
 851          */
 852         if (error == 0) {
 853                 if (ip) {
 854                         --ip->ino_data.nlinks;  /* do before we might block */
 855                         ip->ino_data.ctime = trans->time;
 856                 }
 857                 dip->ino_data.mtime = trans->time;
 858                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 859                 if (ip) {
 860                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 861                         if (ip->ino_data.nlinks == 0 &&
 862                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 863                                 hammer_done_cursor(cursor);
 864                                 hammer_inode_unloadable_check(ip, 1);
 865                                 hammer_flush_inode(ip, 0);
 866                         }
 867                 }
 868
 869         }
 870         return(error);
 871 }
 872
 873 /*
 874  * Add a record to an inode.
 875  *
 876  * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and
 877  * initialize the following additional fields that are not initialized by these
 878  * functions.
 879  *
 880  * The related inode should be share-locked by the caller.  The caller is
 881  * on the frontend.
 882  *
 883  * record->leaf.base.key
 884  * record->leaf.base.rec_type
 885  * record->leaf.base.localization
 886  */
 887 int
 888 hammer_ip_add_record(hammer_transaction_t trans, hammer_record_t record)
 889 {
 890         hammer_inode_t ip = record->ip;
 891         int error;
 892
 893         KKASSERT(record->leaf.base.localization != 0);
 894         record->leaf.base.obj_id = ip->obj_id;
 895         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 896         error = hammer_mem_add(record);
 897         return(error);
 898 }
 899
 900 /*
 901  * Locate a pre-existing bulk record in memory.  The caller wishes to
 902  * replace the record with a new one.  The existing record may have a
 903  * different length (and thus a different key) so we have to use an
 904  * overlap check function.
 905  */
 906 static hammer_record_t
 907 hammer_ip_get_bulk(hammer_record_t record)
 908 {
 909         struct hammer_bulk_info info;
 910         hammer_inode_t ip = record->ip;
 911
 912         info.record = record;
 913         info.conflict = NULL;
 914         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 915                                    hammer_bulk_scan_callback, &info);
 916
 917         return(info.conflict);  /* may be NULL */
 918 }
 919
 920 /*
 921  * Take records vetted by overlap_cmp.  The first non-deleted record
 922  * (if any) stops the scan.
 923  */
 924 static int
 925 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 926 {
 927         struct hammer_bulk_info *info = data;
 928
 929         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 930                              HAMMER_RECF_COMMITTED)) {
 931                 return(0);
 932         }
 933         hammer_ref(&record->lock);
 934         info->conflict = record;
 935         return(-1);                     /* stop scan */
 936 }
 937
 938 /*
 939  * Reserve blockmap space placemarked with an in-memory record.
 940  *
 941  * This routine is called by the frontend in order to be able to directly
 942  * flush a buffer cache buffer.  The frontend has locked the related buffer
 943  * cache buffers and we should be able to manipulate any overlapping
 944  * in-memory records.
 945  *
 946  * The caller is responsible for adding the returned record and deleting
 947  * the returned conflicting record (if any), typically by calling
 948  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
 949  */
 950 hammer_record_t
 951 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 952                    int *errorp)
 953 {
 954         hammer_record_t record;
 955         hammer_dedup_cache_t dcp;
 956         hammer_crc_t crc;
 957         int zone;
 958
 959         /*
 960          * Create a record to cover the direct write.  The record cannot
 961          * be added to the in-memory RB tree here as it might conflict
 962          * with an existing memory record.  See hammer_io_direct_write().
 963          *
 964          * The backend is responsible for finalizing the space reserved in
 965          * this record.
 966          *
 967          * XXX bytes not aligned, depend on the reservation code to
 968          * align the reservation.
 969          */
 970         record = hammer_alloc_mem_record(ip, 0);
 971         zone = hammer_data_zone_index(bytes);
 972         if (bytes == 0)
 973                 crc = 0;
 974         else
 975                 crc = crc32(data, bytes);
 976
 977         if (hammer_live_dedup == 0)
 978                 goto nodedup;
 979         if ((dcp = hammer_dedup_cache_lookup(ip->hmp, crc)) != NULL) {
 980                 struct hammer_dedup_cache tmp = *dcp;
 981
 982                 record->resv = hammer_blockmap_reserve_dedup(ip->hmp, zone,
 983                         bytes, tmp.data_offset, errorp);
 984                 if (record->resv == NULL)
 985                         goto nodedup;
 986
 987                 if (!hammer_dedup_validate(&tmp, zone, bytes, data)) {
 988                         hammer_blockmap_reserve_complete(ip->hmp, record->resv);
 989                         goto nodedup;
 990                 }
 991
 992                 record->leaf.data_offset = tmp.data_offset;
 993                 record->flags |= HAMMER_RECF_DEDUPED;
 994         } else {
 995 nodedup:
 996                 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 997                        &record->leaf.data_offset, errorp);
 998                 if (record->resv == NULL) {
 999                         hdkprintf("reservation failed\n");
1000                         hammer_rel_mem_record(record);
1001                         return(NULL);
1002                 }
1003         }
1004
1005         record->type = HAMMER_MEM_RECORD_DATA;
1006         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
1007         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
1008         record->leaf.base.obj_id = ip->obj_id;
1009         record->leaf.base.key = file_offset + bytes;
1010         record->leaf.base.localization = ip->obj_localization |
1011                                          HAMMER_LOCALIZE_MISC;
1012         record->leaf.data_len = bytes;
1013         record->leaf.data_crc = crc;
1014         KKASSERT(*errorp == 0);
1015
1016         return(record);
1017 }
1018
1019 /*
1020  * Called by hammer_io_direct_write() prior to any possible completion
1021  * of the BIO to emplace the memory record associated with the I/O and
1022  * to replace any prior memory record which might still be active.
1023  *
1024  * Setting the FE deleted flag on the old record (if any) avoids any RB
1025  * tree insertion conflict, amoung other things.
1026  *
1027  * This has to be done prior to the caller completing any related buffer
1028  * cache I/O or a reinstantiation of the buffer may load data from the
1029  * old media location instead of the new media location.  The holding
1030  * of the locked buffer cache buffer serves to interlock the record
1031  * replacement operation.
1032  */
1033 void
1034 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1035 {
1036         hammer_record_t conflict;
1037         int error __debugvar;
1038
1039         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1040                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1041                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1042                         break;
1043                 }
1044                 conflict->flags |= HAMMER_RECF_WANTED;
1045                 tsleep(conflict, 0, "hmrrc3", 0);
1046                 hammer_rel_mem_record(conflict);
1047         }
1048         error = hammer_mem_add(record);
1049         if (conflict)
1050                 hammer_rel_mem_record(conflict);
1051         KKASSERT(error == 0);
1052 }
1053
1054 /*
1055  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1056  * and records in a flushing state are handled by the backend.  The vnops
1057  * setattr code will handle the block containing the truncation point.
1058  *
1059  * Partial blocks are not deleted.
1060  *
1061  * This code is only called on regular files.
1062  */
1063 int
1064 hammer_ip_frontend_trunc(hammer_inode_t ip, off_t file_size)
1065 {
1066         struct rec_trunc_info info;
1067
1068         switch(ip->ino_data.obj_type) {
1069         case HAMMER_OBJTYPE_REGFILE:
1070                 info.rec_type = HAMMER_RECTYPE_DATA;
1071                 break;
1072         case HAMMER_OBJTYPE_DBFILE:
1073                 info.rec_type = HAMMER_RECTYPE_DB;
1074                 break;
1075         default:
1076                 return(EINVAL);
1077         }
1078         info.trunc_off = file_size;
1079         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1080                                    hammer_frontend_trunc_callback, &info);
1081         return(0);
1082 }
1083
1084 /*
1085  * Scan callback for frontend records to destroy during a truncation.
1086  * We must ensure that DELETED_FE is set on the record or the frontend
1087  * will get confused in future read() calls.
1088  *
1089  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1090  *       In this rare case we must wait for the interlock to be cleared.
1091  *
1092  * NOTE: This function is only called on regular files.  There are further
1093  *       restrictions to the setting of DELETED_FE on directory records
1094  *       undergoing a flush due to sensitive inode link count calculations.
1095  */
1096 static int
1097 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1098 {
1099         if (record->flags & HAMMER_RECF_DELETED_FE)
1100                 return(0);
1101 #if 0
1102         if (record->flush_state == HAMMER_FST_FLUSH)
1103                 return(0);
1104 #endif
1105         hammer_ref(&record->lock);
1106         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1107                 hammer_wait_mem_record_ident(record, "hmmtrr");
1108         record->flags |= HAMMER_RECF_DELETED_FE;
1109         hammer_rel_mem_record(record);
1110         return(0);
1111 }
1112
1113 /*
1114  * Return 1 if the caller must check for and delete existing records
1115  * before writing out a new data record.
1116  *
1117  * Return 0 if the caller can just insert the record into the B-Tree without
1118  * checking.
1119  */
1120 static int
1121 hammer_record_needs_overwrite_delete(hammer_record_t record)
1122 {
1123         hammer_inode_t ip = record->ip;
1124         int64_t file_offset;
1125         int r;
1126
1127         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1128                 file_offset = record->leaf.base.key;
1129         else
1130                 file_offset = record->leaf.base.key - record->leaf.data_len;
1131         r = (file_offset < ip->save_trunc_off);
1132         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1133                 if (ip->save_trunc_off <= record->leaf.base.key)
1134                         ip->save_trunc_off = record->leaf.base.key + 1;
1135         } else {
1136                 if (ip->save_trunc_off < record->leaf.base.key)
1137                         ip->save_trunc_off = record->leaf.base.key;
1138         }
1139         return(r);
1140 }
1141
1142 /*
1143  * Backend code.  Sync a record to the media.
1144  */
1145 int
1146 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1147 {
1148         hammer_transaction_t trans = cursor->trans;
1149         int64_t file_offset;
1150         int bytes;
1151         void *bdata;
1152         int error;
1153         int doprop;
1154
1155         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1156         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1157         KKASSERT(record->leaf.base.localization != 0);
1158
1159         /*
1160          * Any direct-write related to the record must complete before we
1161          * can sync the record to the on-disk media.
1162          */
1163         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1164                 hammer_io_direct_wait(record);
1165
1166         /*
1167          * If this is a bulk-data record placemarker there may be an existing
1168          * record on-disk, indicating a data overwrite.  If there is the
1169          * on-disk record must be deleted before we can insert our new record.
1170          *
1171          * We've synthesized this record and do not know what the create_tid
1172          * on-disk is, nor how much data it represents.
1173          *
1174          * Keep in mind that (key) for data records is (base_offset + len),
1175          * not (base_offset).  Also, we only want to get rid of on-disk
1176          * records since we are trying to sync our in-memory record, call
1177          * hammer_ip_delete_range() with truncating set to 1 to make sure
1178          * it skips in-memory records.
1179          *
1180          * It is ok for the lookup to return ENOENT.
1181          *
1182          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1183          * to call hammer_ip_delete_range() or not.  This also means we must
1184          * update sync_trunc_off() as we write.
1185          */
1186         if (record->type == HAMMER_MEM_RECORD_DATA &&
1187             hammer_record_needs_overwrite_delete(record)) {
1188                 file_offset = record->leaf.base.key - record->leaf.data_len;
1189                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1190                         ~HAMMER_BUFMASK;
1191                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1192                 error = hammer_ip_delete_range(
1193                                 cursor, record->ip,
1194                                 file_offset, file_offset + bytes - 1,
1195                                 1);
1196                 if (error && error != ENOENT)
1197                         goto done;
1198         }
1199
1200         /*
1201          * If this is a general record there may be an on-disk version
1202          * that must be deleted before we can insert the new record.
1203          */
1204         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1205                 error = hammer_delete_general(cursor, record->ip, &record->leaf);
1206                 if (error && error != ENOENT)
1207                         goto done;
1208         }
1209
1210         /*
1211          * Setup the cursor.
1212          */
1213         hammer_normalize_cursor(cursor);
1214         cursor->key_beg = record->leaf.base;
1215         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1216         cursor->flags |= HAMMER_CURSOR_BACKEND;
1217         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1218
1219         /*
1220          * Records can wind up on-media before the inode itself is on-media.
1221          * Flag the case.
1222          */
1223         record->ip->flags |= HAMMER_INODE_DONDISK;
1224
1225         /*
1226          * If we are deleting a directory entry an exact match must be
1227          * found on-disk.
1228          */
1229         if (record->type == HAMMER_MEM_RECORD_DEL) {
1230                 error = hammer_btree_lookup(cursor);
1231                 if (error == 0) {
1232                         KKASSERT(cursor->iprec == NULL);
1233                         error = hammer_ip_delete_record(cursor, record->ip,
1234                                                         trans->tid);
1235                         if (error == 0) {
1236                                 record->flags |= HAMMER_RECF_DELETED_BE |
1237                                                  HAMMER_RECF_COMMITTED;
1238                                 ++record->ip->rec_generation;
1239                         }
1240                 }
1241                 goto done;
1242         }
1243
1244         /*
1245          * We are inserting.
1246          *
1247          * Issue a lookup to position the cursor and locate the insertion
1248          * point.  The target key should not exist.  If we are creating a
1249          * directory entry we may have to iterate the low 32 bits of the
1250          * key to find an unused key.
1251          */
1252         hammer_sync_lock_sh(trans);
1253         cursor->flags |= HAMMER_CURSOR_INSERT;
1254         error = hammer_btree_lookup(cursor);
1255         if (hammer_debug_inode)
1256                 hdkprintf("DOINSERT LOOKUP %d\n", error);
1257         if (error == 0) {
1258                 hdkprintf("duplicate rec at (%016jx)\n",
1259                         (intmax_t)record->leaf.base.key);
1260                 if (hammer_debug_critical)
1261                         Debugger("duplicate record1");
1262                 error = EIO;
1263         }
1264
1265         if (error != ENOENT)
1266                 goto done_unlock;
1267
1268         /*
1269          * Allocate the record and data.  The result buffers will be
1270          * marked as being modified and further calls to
1271          * hammer_modify_buffer() will result in unneeded UNDO records.
1272          *
1273          * Support zero-fill records (data == NULL and data_len != 0)
1274          */
1275         if (record->type == HAMMER_MEM_RECORD_DATA) {
1276                 /*
1277                  * The data portion of a bulk-data record has already been
1278                  * committed to disk, we need only adjust the layer2
1279                  * statistics in the same transaction as our B-Tree insert.
1280                  */
1281                 KKASSERT(record->leaf.data_offset != 0);
1282                 error = hammer_blockmap_finalize(trans,
1283                                                  record->resv,
1284                                                  record->leaf.data_offset,
1285                                                  record->leaf.data_len);
1286
1287                 if (hammer_live_dedup == 2 &&
1288                     (record->flags & HAMMER_RECF_DEDUPED) == 0) {
1289                         hammer_dedup_cache_add(record->ip, &record->leaf);
1290                 }
1291         } else if (record->data && record->leaf.data_len) {
1292                 /*
1293                  * Wholely cached record, with data.  Allocate the data.
1294                  */
1295                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1296                                           record->leaf.base.rec_type,
1297                                           &record->leaf.data_offset,
1298                                           &cursor->data_buffer,
1299                                           0, &error);
1300                 if (bdata == NULL)
1301                         goto done_unlock;
1302                 hammer_crc_set_leaf(record->data, &record->leaf);
1303                 hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1304                 bcopy(record->data, bdata, record->leaf.data_len);
1305                 hammer_modify_buffer_done(cursor->data_buffer);
1306         } else {
1307                 /*
1308                  * Wholely cached record, without data.
1309                  */
1310                 record->leaf.data_offset = 0;
1311                 record->leaf.data_crc = 0;
1312         }
1313
1314         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1315         if (hammer_debug_inode && error) {
1316                 hdkprintf("BTREE INSERT error %d @ %016jx:%d key %016jx\n",
1317                         error,
1318                         (intmax_t)cursor->node->node_offset,
1319                         cursor->index,
1320                         (intmax_t)record->leaf.base.key);
1321         }
1322
1323         /*
1324          * Our record is on-disk and we normally mark the in-memory version
1325          * as having been committed (and not BE-deleted).
1326          *
1327          * If the record represented a directory deletion but we had to
1328          * sync a valid directory entry to disk due to dependancies,
1329          * we must convert the record to a covering delete so the
1330          * frontend does not have visibility on the synced entry.
1331          *
1332          * WARNING: cursor's leaf pointer may have changed after do_propagation
1333          *          returns!
1334          */
1335         if (error == 0) {
1336                 if (doprop) {
1337                         hammer_btree_do_propagation(cursor, &record->leaf);
1338                 }
1339                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1340                         /*
1341                          * Must convert deleted directory entry add
1342                          * to a directory entry delete.
1343                          */
1344                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1345                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1346                         record->type = HAMMER_MEM_RECORD_DEL;
1347                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1348                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1349                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1350                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1351                                                  HAMMER_RECF_DELETED_BE)) == 0);
1352                         /* converted record is not yet committed */
1353                         /* hammer_flush_record_done takes care of the rest */
1354                 } else {
1355                         /*
1356                          * Everything went fine and we are now done with
1357                          * this record.
1358                          */
1359                         record->flags |= HAMMER_RECF_COMMITTED;
1360                         ++record->ip->rec_generation;
1361                 }
1362         } else {
1363                 if (record->leaf.data_offset) {
1364                         hammer_blockmap_free(trans, record->leaf.data_offset,
1365                                              record->leaf.data_len);
1366                 }
1367         }
1368 done_unlock:
1369         hammer_sync_unlock(trans);
1370 done:
1371         return(error);
1372 }
1373
1374 /*
1375  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1376  * entry's key is used to deal with hash collisions in the upper 32 bits.
1377  * A unique 64 bit key is generated in-memory and may be regenerated a
1378  * second time when the directory record is flushed to the on-disk B-Tree.
1379  *
1380  * A referenced record is passed to this function.  This function
1381  * eats the reference.  If an error occurs the record will be deleted.
1382  *
1383  * A copy of the temporary record->data pointer provided by the caller
1384  * will be made.
1385  */
1386 int
1387 hammer_mem_add(hammer_record_t record)
1388 {
1389         hammer_mount_t hmp = record->ip->hmp;
1390
1391         /*
1392          * Make a private copy of record->data
1393          */
1394         if (record->data)
1395                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1396
1397         /*
1398          * Insert into the RB tree.  A unique key should have already
1399          * been selected if this is a directory entry.
1400          */
1401         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1402                 record->flags |= HAMMER_RECF_DELETED_FE;
1403                 hammer_rel_mem_record(record);
1404                 return (EEXIST);
1405         }
1406         ++hmp->rsv_recs;
1407         ++record->ip->rsv_recs;
1408         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1409         record->flags |= HAMMER_RECF_ONRBTREE;
1410         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1411         hammer_rel_mem_record(record);
1412         return(0);
1413 }
1414
1415 /************************************************************************
1416  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1417  ************************************************************************
1418  *
1419  * These functions augment the B-Tree scanning functions in hammer_btree.c
1420  * by merging in-memory records with on-disk records.
1421  */
1422
1423 /*
1424  * Locate a particular record either in-memory or on-disk.
1425  *
1426  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1427  * NOT be called to iterate results.
1428  */
1429 int
1430 hammer_ip_lookup(hammer_cursor_t cursor)
1431 {
1432         int error;
1433
1434         /*
1435          * If the element is in-memory return it without searching the
1436          * on-disk B-Tree
1437          */
1438         KKASSERT(cursor->ip);
1439         error = hammer_mem_lookup(cursor);
1440         if (error == 0) {
1441                 cursor->leaf = &cursor->iprec->leaf;
1442                 return(error);
1443         }
1444         if (error != ENOENT)
1445                 return(error);
1446
1447         /*
1448          * If the inode has on-disk components search the on-disk B-Tree.
1449          */
1450         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1451                 return(error);
1452         error = hammer_btree_lookup(cursor);
1453         if (error == 0)
1454                 error = hammer_btree_extract_leaf(cursor);
1455         return(error);
1456 }
1457
1458 /*
1459  * Helper for hammer_ip_first()/hammer_ip_next()
1460  *
1461  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1462  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1463  * state if hammer_ip_next() needs to re-seek.
1464  */
1465 static __inline
1466 int
1467 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1468 {
1469         hammer_inode_t ip = cursor->ip;
1470         int error;
1471
1472         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1473                 error = hammer_btree_lookup(cursor);
1474                 if (error == ENOENT || error == EDEADLK) {
1475                         if (hammer_debug_general & 0x2000) {
1476                                 hdkprintf("error %d node %p %016jx index %d\n",
1477                                         error, cursor->node,
1478                                         (intmax_t)cursor->node->node_offset,
1479                                         cursor->index);
1480                         }
1481                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1482                         error = hammer_btree_iterate(cursor);
1483                 }
1484                 if (error == 0) {
1485                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1486                                            HAMMER_CURSOR_ATEDISK);
1487                 } else {
1488                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1489                                          HAMMER_CURSOR_ATEDISK;
1490                         if (error == ENOENT)
1491                                 error = 0;
1492                 }
1493         } else {
1494                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1495                 error = 0;
1496         }
1497         return(error);
1498 }
1499
1500 /*
1501  * Helper for hammer_ip_next()
1502  *
1503  * The caller has determined that the media cursor is further along than the
1504  * memory cursor and must be reseeked after a generation number change.
1505  */
1506 static
1507 int
1508 _hammer_ip_reseek(hammer_cursor_t cursor)
1509 {
1510         struct hammer_base_elm save;
1511         hammer_btree_elm_t elm;
1512         int error __debugvar;
1513         int r;
1514         int again = 0;
1515
1516         /*
1517          * Do the re-seek.
1518          */
1519         hkprintf("Debug: re-seeked during scan @ino=%016jx\n",
1520                 (intmax_t)cursor->ip->obj_id);
1521         save = cursor->key_beg;
1522         cursor->key_beg = cursor->iprec->leaf.base;
1523         error = _hammer_ip_seek_btree(cursor);
1524         KKASSERT(error == 0);
1525         cursor->key_beg = save;
1526
1527         /*
1528          * If the memory record was previous returned to
1529          * the caller and the media record matches
1530          * (-1/+1: only create_tid differs), then iterate
1531          * the media record to avoid a double result.
1532          */
1533         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1534             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1535                 elm = &cursor->node->ondisk->elms[cursor->index];
1536                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1537                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1538                         if (r >= -1 && r <= 1) {
1539                                 hkprintf("Debug: iterated after "
1540                                         "re-seek (asof r=%d)\n", r);
1541                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1542                                 again = 1;
1543                         }
1544                 } else {
1545                         if (r == 0) {
1546                                 hkprintf("Debug: iterated after "
1547                                         "re-seek\n");
1548                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1549                                 again = 1;
1550                         }
1551                 }
1552         }
1553         return(again);
1554 }
1555
1556 /*
1557  * Locate the first record within the cursor's key_beg/key_end range,
1558  * restricted to a particular inode.  0 is returned on success, ENOENT
1559  * if no records matched the requested range, or some other error.
1560  *
1561  * When 0 is returned hammer_ip_next() may be used to iterate additional
1562  * records within the requested range.
1563  *
1564  * This function can return EDEADLK, requiring the caller to terminate
1565  * the cursor and try again.
1566  */
1567
1568 int
1569 hammer_ip_first(hammer_cursor_t cursor)
1570 {
1571         hammer_inode_t ip __debugvar = cursor->ip;
1572         int error;
1573
1574         KKASSERT(ip != NULL);
1575
1576         /*
1577          * Clean up fields and setup for merged scan
1578          */
1579         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1580
1581         /*
1582          * Search the in-memory record list (Red-Black tree).  Unlike the
1583          * B-Tree search, mem_first checks for records in the range.
1584          *
1585          * This function will setup both ATEMEM and MEMEOF properly for
1586          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1587          */
1588         hammer_mem_first(cursor);
1589
1590         /*
1591          * Detect generation changes during blockages, including
1592          * blockages which occur on the initial btree search.
1593          */
1594         cursor->rec_generation = cursor->ip->rec_generation;
1595
1596         /*
1597          * Initial search and result
1598          */
1599         error = _hammer_ip_seek_btree(cursor);
1600         if (error == 0)
1601                 error = hammer_ip_next(cursor);
1602
1603         return (error);
1604 }
1605
1606 /*
1607  * Retrieve the next record in a merged iteration within the bounds of the
1608  * cursor.  This call may be made multiple times after the cursor has been
1609  * initially searched with hammer_ip_first().
1610  *
1611  * There are numerous special cases in this code to deal with races between
1612  * in-memory records and on-media records.
1613  *
1614  * 0 is returned on success, ENOENT if no further records match the
1615  * requested range, or some other error code is returned.
1616  */
1617 int
1618 hammer_ip_next(hammer_cursor_t cursor)
1619 {
1620         hammer_btree_elm_t elm;
1621         hammer_record_t rec;
1622         hammer_record_t tmprec;
1623         int error;
1624         int r;
1625
1626 again:
1627         /*
1628          * Get the next on-disk record
1629          *
1630          * NOTE: If we deleted the last on-disk record we had scanned
1631          *       ATEDISK will be clear and RETEST will be set, forcing
1632          *       a call to iterate.  The fact that ATEDISK is clear causes
1633          *       iterate to re-test the 'current' element.  If ATEDISK is
1634          *       set, iterate will skip the 'current' element.
1635          */
1636         error = 0;
1637         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1638                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1639                                      HAMMER_CURSOR_RETEST)) {
1640                         error = hammer_btree_iterate(cursor);
1641                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1642                         if (error == 0) {
1643                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1644                                 hammer_cache_node(&cursor->ip->cache[1],
1645                                                   cursor->node);
1646                         } else if (error == ENOENT) {
1647                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1648                                                  HAMMER_CURSOR_ATEDISK;
1649                                 error = 0;
1650                         }
1651                 }
1652         }
1653
1654         /*
1655          * If the generation changed the backend has deleted or committed
1656          * one or more memory records since our last check.
1657          *
1658          * When this case occurs if the disk cursor is > current memory record
1659          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1660          * Since the cursor is ahead it must have not yet been eaten (if
1661          * not at eof anyway). (XXX data offset case?)
1662          *
1663          * NOTE: we are not doing a full check here.  That will be handled
1664          * later on.
1665          *
1666          * If we have exhausted all memory records we do not have to do any
1667          * further seeks.
1668          */
1669         while (cursor->rec_generation != cursor->ip->rec_generation &&
1670                error == 0) {
1671                 hkprintf("Debug: generation changed during scan @ino=%016jx\n",
1672                         (intmax_t)cursor->ip->obj_id);
1673                 cursor->rec_generation = cursor->ip->rec_generation;
1674                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1675                         break;
1676                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1677                         r = 1;
1678                 } else {
1679                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1680                         elm = &cursor->node->ondisk->elms[cursor->index];
1681                         r = hammer_btree_cmp(&elm->base,
1682                                              &cursor->iprec->leaf.base);
1683                 }
1684
1685                 /*
1686                  * Do we re-seek the media cursor?
1687                  */
1688                 if (r > 0) {
1689                         if (_hammer_ip_reseek(cursor))
1690                                 goto again;
1691                 }
1692         }
1693
1694         /*
1695          * We can now safely get the next in-memory record.  We cannot
1696          * block here.
1697          *
1698          * hammer_rec_scan_cmp:  Is the record still in our general range,
1699          *                       (non-inclusive of snapshot exclusions)?
1700          * hammer_rec_scan_callback: Is the record in our snapshot?
1701          */
1702         tmprec = NULL;
1703         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1704                 /*
1705                  * If the current memory record was eaten then get the next
1706                  * one.  Stale records are skipped.
1707                  */
1708                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1709                         tmprec = cursor->iprec;
1710                         cursor->iprec = NULL;
1711                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1712                         while (rec) {
1713                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1714                                         break;
1715                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1716                                         break;
1717                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1718                         }
1719                         if (cursor->iprec) {
1720                                 KKASSERT(cursor->iprec == rec);
1721                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1722                         } else {
1723                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1724                         }
1725                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1726                 }
1727         }
1728
1729         /*
1730          * MEMORY RECORD VALIDITY TEST
1731          *
1732          * (We still can't block, which is why tmprec is being held so
1733          * long).
1734          *
1735          * If the memory record is no longer valid we skip it.  It may
1736          * have been deleted by the frontend.  If it was deleted or
1737          * committed by the backend the generation change re-seeked the
1738          * disk cursor and the record will be present there.
1739          */
1740         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1741                 KKASSERT(cursor->iprec);
1742                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1743                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1744                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1745                         if (tmprec)
1746                                 hammer_rel_mem_record(tmprec);
1747                         goto again;
1748                 }
1749         }
1750         if (tmprec)
1751                 hammer_rel_mem_record(tmprec);
1752
1753         /*
1754          * Extract either the disk or memory record depending on their
1755          * relative position.
1756          */
1757         error = 0;
1758         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1759         case 0:
1760                 /*
1761                  * Both entries valid.   Compare the entries and nominally
1762                  * return the first one in the sort order.  Numerous cases
1763                  * require special attention, however.
1764                  */
1765                 elm = &cursor->node->ondisk->elms[cursor->index];
1766                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1767
1768                 /*
1769                  * If the two entries differ only by their key (-2/2) or
1770                  * create_tid (-1/1), and are DATA records, we may have a
1771                  * nominal match.  We have to calculate the base file
1772                  * offset of the data.
1773                  */
1774                 if (r <= 2 && r >= -2 && r != 0 &&
1775                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1776                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1777                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1778                         int64_t base2 = cursor->iprec->leaf.base.key -
1779                                         cursor->iprec->leaf.data_len;
1780                         if (base1 == base2)
1781                                 r = 0;
1782                 }
1783
1784                 if (r < 0) {
1785                         error = hammer_btree_extract_leaf(cursor);
1786                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1787                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1788                         break;
1789                 }
1790
1791                 /*
1792                  * If the entries match exactly the memory entry is either
1793                  * an on-disk directory entry deletion or a bulk data
1794                  * overwrite.  If it is a directory entry deletion we eat
1795                  * both entries.
1796                  *
1797                  * For the bulk-data overwrite case it is possible to have
1798                  * visibility into both, which simply means the syncer
1799                  * hasn't gotten around to doing the delete+insert sequence
1800                  * on the B-Tree.  Use the memory entry and throw away the
1801                  * on-disk entry.
1802                  *
1803                  * If the in-memory record is not either of these we
1804                  * probably caught the syncer while it was syncing it to
1805                  * the media.  Since we hold a shared lock on the cursor,
1806                  * the in-memory record had better be marked deleted at
1807                  * this point.
1808                  */
1809                 if (r == 0) {
1810                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1811                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1812                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1813                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1814                                         goto again;
1815                                 }
1816                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1817                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1818                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1819                                 }
1820                                 /* fall through to memory entry */
1821                         } else {
1822                                 hpanic("duplicate mem/B-Tree entry %p %d %08x",
1823                                         cursor->iprec,
1824                                         cursor->iprec->type,
1825                                         cursor->iprec->flags);
1826                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1827                                 goto again;
1828                         }
1829                 }
1830                 /* fall through to the memory entry */
1831         case HAMMER_CURSOR_ATEDISK:
1832                 /*
1833                  * Only the memory entry is valid.
1834                  */
1835                 cursor->leaf = &cursor->iprec->leaf;
1836                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1837                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1838
1839                 /*
1840                  * If the memory entry is an on-disk deletion we should have
1841                  * also had found a B-Tree record.  If the backend beat us
1842                  * to it it would have interlocked the cursor and we should
1843                  * have seen the in-memory record marked DELETED_FE.
1844                  */
1845                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1846                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1847                         hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x",
1848                                 cursor->iprec,
1849                                 cursor->iprec->flags);
1850                 }
1851                 break;
1852         case HAMMER_CURSOR_ATEMEM:
1853                 /*
1854                  * Only the disk entry is valid
1855                  */
1856                 error = hammer_btree_extract_leaf(cursor);
1857                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1858                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1859                 break;
1860         default:
1861                 /*
1862                  * Neither entry is valid
1863                  *
1864                  * XXX error not set properly
1865                  */
1866                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1867                 cursor->leaf = NULL;
1868                 error = ENOENT;
1869                 break;
1870         }
1871         return(error);
1872 }
1873
1874 /*
1875  * Resolve the cursor->data pointer for the current cursor position in
1876  * a merged iteration.
1877  */
1878 int
1879 hammer_ip_resolve_data(hammer_cursor_t cursor)
1880 {
1881         hammer_record_t record;
1882         int error;
1883
1884         if (hammer_cursor_inmem(cursor)) {
1885                 /*
1886                  * The data associated with an in-memory record is usually
1887                  * kmalloced, but reserve-ahead data records will have an
1888                  * on-disk reference.
1889                  *
1890                  * NOTE: Reserve-ahead data records must be handled in the
1891                  *       context of the related high level buffer cache buffer
1892                  *       to interlock against async writes.
1893                  *
1894                  * NOTE: We might catch a direct write in-progress, in which
1895                  *       case we must wait for it to complete.  The wait
1896                  *       function will also clean out any buffer aliases.
1897                  *
1898                  *       (In fact, it is possible that the write had not
1899                  *        even started yet).
1900                  */
1901                 record = cursor->iprec;
1902                 cursor->data = record->data;
1903                 error = 0;
1904                 if (cursor->data == NULL) {
1905                         hammer_io_direct_wait(record);
1906                         KKASSERT(record->leaf.base.rec_type ==
1907                                  HAMMER_RECTYPE_DATA);
1908                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1909                                                     record->leaf.data_offset,
1910                                                     record->leaf.data_len,
1911                                                     &error,
1912                                                     &cursor->data_buffer);
1913                 }
1914         } else {
1915                 /*
1916                  * Loading leaf here isn't necessary if it's guaranteed that
1917                  * the cursor is at a leaf node (which basically should be)
1918                  * because hammer_btree_extract_data() does that.
1919                  */
1920                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1921                 error = hammer_btree_extract_data(cursor);
1922         }
1923         return(error);
1924 }
1925
1926 /*
1927  * Backend truncation / record replacement - delete records in range.
1928  *
1929  * Delete all records within the specified range for inode ip.  In-memory
1930  * records still associated with the frontend are ignored.
1931  *
1932  * If truncating is non-zero in-memory records associated with the back-end
1933  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1934  *
1935  * NOTES:
1936  *
1937  *      * An unaligned range will cause new records to be added to cover
1938  *        the edge cases. (XXX not implemented yet).
1939  *
1940  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1941  *        also do not deal with unaligned ranges.
1942  *
1943  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1944  *
1945  *      * Record keys for regular file data have to be special-cased since
1946  *        they indicate the end of the range (key = base + bytes).
1947  *
1948  *      * This function may be asked to delete ridiculously huge ranges, for
1949  *        example if someone truncates or removes a 1TB regular file.  We
1950  *        must be very careful on restarts and we may have to stop w/
1951  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1952  */
1953 int
1954 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1955                        int64_t ran_beg, int64_t ran_end, int truncating)
1956 {
1957         hammer_transaction_t trans = cursor->trans;
1958         hammer_btree_leaf_elm_t leaf;
1959         int error;
1960         int64_t off;
1961         int64_t tmp64;
1962
1963         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1964 retry:
1965         hammer_normalize_cursor(cursor);
1966         cursor->key_beg.localization = ip->obj_localization |
1967                                        HAMMER_LOCALIZE_MISC;
1968         cursor->key_beg.obj_id = ip->obj_id;
1969         cursor->key_beg.create_tid = 0;
1970         cursor->key_beg.delete_tid = 0;
1971         cursor->key_beg.obj_type = 0;
1972
1973         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1974                 cursor->key_beg.key = ran_beg;
1975                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1976         } else {
1977                 /*
1978                  * The key in the B-Tree is (base+bytes), so the first possible
1979                  * matching key is ran_beg + 1.
1980                  */
1981                 cursor->key_beg.key = ran_beg + 1;
1982                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1983         }
1984
1985         cursor->key_end = cursor->key_beg;
1986         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1987                 cursor->key_end.key = ran_end;
1988         } else {
1989                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1990                 if (tmp64 < ran_end)
1991                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1992                 else
1993                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1994         }
1995
1996         cursor->asof = ip->obj_asof;
1997         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1998         cursor->flags |= HAMMER_CURSOR_ASOF;
1999         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2000         cursor->flags |= HAMMER_CURSOR_BACKEND;
2001         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
2002
2003         error = hammer_ip_first(cursor);
2004
2005         /*
2006          * Iterate through matching records and mark them as deleted.
2007          */
2008         while (error == 0) {
2009                 leaf = cursor->leaf;
2010
2011                 KKASSERT(leaf->base.delete_tid == 0);
2012                 KKASSERT(leaf->base.obj_id == ip->obj_id);
2013
2014                 /*
2015                  * There may be overlap cases for regular file data.  Also
2016                  * remember the key for a regular file record is (base + len),
2017                  * NOT (base).
2018                  *
2019                  * Note that due to duplicates (mem & media) allowed by
2020                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
2021                  */
2022                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
2023                         off = leaf->base.key - leaf->data_len;
2024                         /*
2025                          * Check the left edge case.  We currently do not
2026                          * split existing records.
2027                          */
2028                         if (off < ran_beg && leaf->base.key > ran_beg) {
2029                                 hpanic("hammer left edge case %016jx %d",
2030                                         (intmax_t)leaf->base.key,
2031                                         leaf->data_len);
2032                         }
2033
2034                         /*
2035                          * Check the right edge case.  Note that the
2036                          * record can be completely out of bounds, which
2037                          * terminates the search.
2038                          *
2039                          * base->key is exclusive of the right edge while
2040                          * ran_end is inclusive of the right edge.  The
2041                          * (key - data_len) left boundary is inclusive.
2042                          *
2043                          * XXX theory-check this test at some point, are
2044                          * we missing a + 1 somewhere?  Note that ran_end
2045                          * could overflow.
2046                          */
2047                         if (leaf->base.key - 1 > ran_end) {
2048                                 if (leaf->base.key - leaf->data_len > ran_end)
2049                                         break;
2050                                 hpanic("hammer right edge case");
2051                         }
2052                 } else {
2053                         off = leaf->base.key;
2054                 }
2055
2056                 /*
2057                  * Delete the record.  When truncating we do not delete
2058                  * in-memory (data) records because they represent data
2059                  * written after the truncation.
2060                  *
2061                  * This will also physically destroy the B-Tree entry and
2062                  * data if the retention policy dictates.  The function
2063                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2064                  * to retest the new 'current' element.
2065                  */
2066                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2067                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2068                         /*
2069                          * If we have built up too many meta-buffers we risk
2070                          * deadlocking the kernel and must stop.  This can
2071                          * occur when deleting ridiculously huge files.
2072                          * sync_trunc_off is updated so the next cycle does
2073                          * not re-iterate records we have already deleted.
2074                          *
2075                          * This is only done with formal truncations.
2076                          */
2077                         if (truncating > 1 && error == 0 &&
2078                             hammer_flusher_meta_limit(ip->hmp)) {
2079                                 ip->sync_trunc_off = off;
2080                                 error = EWOULDBLOCK;
2081                         }
2082                 }
2083                 if (error)
2084                         break;
2085                 ran_beg = off;  /* for restart */
2086                 error = hammer_ip_next(cursor);
2087         }
2088         if (cursor->node)
2089                 hammer_cache_node(&ip->cache[1], cursor->node);
2090
2091         if (error == EDEADLK) {
2092                 hammer_done_cursor(cursor);
2093                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2094                 if (error == 0)
2095                         goto retry;
2096         }
2097         if (error == ENOENT)
2098                 error = 0;
2099         return(error);
2100 }
2101
2102 /*
2103  * This backend function deletes the specified record on-disk, similar to
2104  * delete_range but for a specific record.  Unlike the exact deletions
2105  * used when deleting a directory entry this function uses an ASOF search
2106  * like delete_range.
2107  *
2108  * This function may be called with ip->obj_asof set for a slave snapshot,
2109  * so don't use it.  We always delete non-historical records only.
2110  */
2111 static int
2112 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2113                       hammer_btree_leaf_elm_t leaf)
2114 {
2115         hammer_transaction_t trans = cursor->trans;
2116         int error;
2117
2118         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2119 retry:
2120         hammer_normalize_cursor(cursor);
2121         cursor->key_beg = leaf->base;
2122         cursor->asof = HAMMER_MAX_TID;
2123         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2124         cursor->flags |= HAMMER_CURSOR_ASOF;
2125         cursor->flags |= HAMMER_CURSOR_BACKEND;
2126         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2127
2128         error = hammer_btree_lookup(cursor);
2129         if (error == 0) {
2130                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2131         }
2132         if (error == EDEADLK) {
2133                 hammer_done_cursor(cursor);
2134                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2135                 if (error == 0)
2136                         goto retry;
2137         }
2138         return(error);
2139 }
2140
2141 /*
2142  * This function deletes remaining auxillary records when an inode is
2143  * being deleted.  This function explicitly does not delete the
2144  * inode record, directory entry, data, or db records.  Those must be
2145  * properly disposed of prior to this call.
2146  */
2147 int
2148 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2149 {
2150         hammer_transaction_t trans = cursor->trans;
2151         hammer_btree_leaf_elm_t leaf __debugvar;
2152         int error;
2153
2154         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2155 retry:
2156         hammer_normalize_cursor(cursor);
2157         cursor->key_beg.localization = ip->obj_localization |
2158                                        HAMMER_LOCALIZE_MISC;
2159         cursor->key_beg.obj_id = ip->obj_id;
2160         cursor->key_beg.create_tid = 0;
2161         cursor->key_beg.delete_tid = 0;
2162         cursor->key_beg.obj_type = 0;
2163         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2164         cursor->key_beg.key = HAMMER_MIN_KEY;
2165
2166         cursor->key_end = cursor->key_beg;
2167         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2168         cursor->key_end.key = HAMMER_MAX_KEY;
2169
2170         cursor->asof = ip->obj_asof;
2171         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2172         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2173         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2174         cursor->flags |= HAMMER_CURSOR_BACKEND;
2175
2176         error = hammer_ip_first(cursor);
2177
2178         /*
2179          * Iterate through matching records and mark them as deleted.
2180          */
2181         while (error == 0) {
2182                 leaf = cursor->leaf;
2183
2184                 KKASSERT(leaf->base.delete_tid == 0);
2185
2186                 /*
2187                  * Mark the record and B-Tree entry as deleted.  This will
2188                  * also physically delete the B-Tree entry, record, and
2189                  * data if the retention policy dictates.  The function
2190                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2191                  * to retest the new 'current' element.
2192                  *
2193                  * Directory entries (and delete-on-disk directory entries)
2194                  * must be synced and cannot be deleted.
2195                  */
2196                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2197                 ++*countp;
2198                 if (error)
2199                         break;
2200                 error = hammer_ip_next(cursor);
2201         }
2202         if (cursor->node)
2203                 hammer_cache_node(&ip->cache[1], cursor->node);
2204         if (error == EDEADLK) {
2205                 hammer_done_cursor(cursor);
2206                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2207                 if (error == 0)
2208                         goto retry;
2209         }
2210         if (error == ENOENT)
2211                 error = 0;
2212         return(error);
2213 }
2214
2215 /*
2216  * Delete the record at the current cursor.  On success the cursor will
2217  * be positioned appropriately for an iteration but may no longer be at
2218  * a leaf node.
2219  *
2220  * This routine is only called from the backend.
2221  *
2222  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2223  * cursor and retry.
2224  */
2225 int
2226 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2227                         hammer_tid_t tid)
2228 {
2229         hammer_record_t iprec;
2230         int error;
2231
2232         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2233         KKASSERT(tid != 0);
2234
2235         /*
2236          * In-memory (unsynchronized) records can simply be freed.  This
2237          * only occurs in range iterations since all other records are
2238          * individually synchronized.  Thus there should be no confusion with
2239          * the interlock.
2240          *
2241          * An in-memory record may be deleted before being committed to disk,
2242          * but could have been accessed in the mean time.  The reservation
2243          * code will deal with the case.
2244          */
2245         if (hammer_cursor_inmem(cursor)) {
2246                 iprec = cursor->iprec;
2247                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2248                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2249                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2250                 KKASSERT(iprec->ip == ip);
2251                 ++ip->rec_generation;
2252                 return(0);
2253         }
2254
2255         /*
2256          * On-disk records are marked as deleted by updating their delete_tid.
2257          * This does not effect their position in the B-Tree (which is based
2258          * on their create_tid).
2259          *
2260          * Frontend B-Tree operations track inodes so we tell
2261          * hammer_delete_at_cursor() not to.
2262          */
2263         error = hammer_btree_extract_leaf(cursor);
2264         if (error == 0) {
2265                 error = hammer_delete_at_cursor(
2266                                 cursor,
2267                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2268                                 cursor->trans->tid,
2269                                 cursor->trans->time32,
2270                                 0, NULL);
2271         }
2272         return(error);
2273 }
2274
2275 /*
2276  * Used to write a generic record w/optional data to the media b-tree
2277  * when no inode context is available.  Used by the mirroring and
2278  * snapshot code.
2279  *
2280  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2281  * flagged for backend operation and not flagged ASOF (since we are
2282  * doing an insertion).
2283  *
2284  * This function will acquire the appropriate sync lock and will set
2285  * the cursor insertion flag for the operation, do the btree lookup,
2286  * and the insertion, and clear the insertion flag and sync lock before
2287  * returning.  The cursor state will be such that the caller can continue
2288  * scanning (used by the mirroring code).
2289  *
2290  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2291  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2292  *
2293  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2294  *                retry.
2295  *
2296  *       EALREADY can be returned if the record already exists (WARNING,
2297  *                because ASOF cannot be used no check is made for illegal
2298  *                duplicates).
2299  *
2300  * NOTE: Do not use the function for normal inode-related records as this
2301  *       functions goes directly to the media and is not integrated with
2302  *       in-memory records.
2303  */
2304 int
2305 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2306                         void *udata, int mode)
2307 {
2308         hammer_transaction_t trans;
2309         hammer_buffer_t data_buffer;
2310         hammer_off_t ndata_offset;
2311         hammer_tid_t high_tid;
2312         void *ndata;
2313         int error;
2314         int doprop;
2315
2316         trans = cursor->trans;
2317         data_buffer = NULL;
2318         ndata_offset = 0;
2319         doprop = 0;
2320
2321         KKASSERT((cursor->flags &
2322                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2323                   (HAMMER_CURSOR_BACKEND));
2324
2325         hammer_sync_lock_sh(trans);
2326
2327         if (leaf->data_len) {
2328                 ndata = hammer_alloc_data(trans, leaf->data_len,
2329                                           leaf->base.rec_type,
2330                                           &ndata_offset, &data_buffer,
2331                                           0, &error);
2332                 if (ndata == NULL) {
2333                         hammer_sync_unlock(trans);
2334                         return (error);
2335                 }
2336                 leaf->data_offset = ndata_offset;
2337                 hammer_modify_buffer_noundo(trans, data_buffer);
2338
2339                 switch(mode) {
2340                 case HAMMER_CREATE_MODE_UMIRROR:
2341                         error = copyin(udata, ndata, leaf->data_len);
2342                         if (error == 0) {
2343                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2344                                         hdkprintf("CRC DATA @ %016jx/%d MISMATCH ON PIPE\n",
2345                                                 (intmax_t)ndata_offset,
2346                                                 leaf->data_len);
2347                                         error = EINVAL;
2348                                 } else {
2349                                         error = hammer_cursor_localize_data(
2350                                                         ndata, leaf);
2351                                 }
2352                         }
2353                         break;
2354                 case HAMMER_CREATE_MODE_SYS:
2355                         bcopy(udata, ndata, leaf->data_len);
2356                         error = 0;
2357                         hammer_crc_set_leaf(ndata, leaf);
2358                         break;
2359                 default:
2360                         hpanic("bad mode %d", mode);
2361                         break; /* NOT REACHED */
2362                 }
2363                 hammer_modify_buffer_done(data_buffer);
2364         } else {
2365                 leaf->data_offset = 0;
2366                 error = 0;
2367                 ndata = NULL;
2368         }
2369         if (error)
2370                 goto failed;
2371
2372         /*
2373          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2374          */
2375         cursor->flags |= HAMMER_CURSOR_INSERT;
2376         error = hammer_btree_lookup(cursor);
2377         if (error != ENOENT) {
2378                 if (error == 0)
2379                         error = EALREADY;
2380                 goto failed;
2381         }
2382         error = hammer_btree_insert(cursor, leaf, &doprop);
2383
2384         /*
2385          * Cursor is left on current element, we want to skip it now.
2386          * (in case the caller is scanning)
2387          */
2388         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2389         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2390
2391         /*
2392          * If the insertion happens to be creating (and not just replacing)
2393          * an inode we have to track it.
2394          */
2395         if (error == 0 &&
2396             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2397             leaf->base.delete_tid == 0) {
2398                 hammer_modify_volume_field(trans, trans->rootvol,
2399                                            vol0_stat_inodes);
2400                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2401                 hammer_modify_volume_done(trans->rootvol);
2402         }
2403
2404         /*
2405          * vol0_next_tid must track the highest TID stored in the filesystem.
2406          * We do not need to generate undo for this update.
2407          */
2408         high_tid = leaf->base.create_tid;
2409         if (high_tid < leaf->base.delete_tid)
2410                 high_tid = leaf->base.delete_tid;
2411         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2412                 hammer_modify_volume_noundo(trans, trans->rootvol);
2413                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2414                 hammer_modify_volume_done(trans->rootvol);
2415         }
2416
2417         /*
2418          * WARNING!  cursor's leaf pointer may have changed after
2419          *           do_propagation returns.
2420          */
2421         if (error == 0 && doprop)
2422                 hammer_btree_do_propagation(cursor, leaf);
2423
2424 failed:
2425         /*
2426          * Cleanup
2427          */
2428         if (error && leaf->data_offset) {
2429                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2430
2431         }
2432         hammer_sync_unlock(trans);
2433         if (data_buffer)
2434                 hammer_rel_buffer(data_buffer, 0);
2435         return (error);
2436 }
2437
2438 /*
2439  * Delete the B-Tree element at the current cursor and do any necessary
2440  * mirror propagation.
2441  *
2442  * The cursor must be properly positioned for an iteration on return but
2443  * may be pointing at an internal element.
2444  *
2445  * An element can be un-deleted by passing a delete_tid of 0 with
2446  * HAMMER_DELETE_ADJUST.
2447  *
2448  * This function will store the number of bytes deleted in *stat_bytes
2449  * if stat_bytes is not NULL.
2450  */
2451 int
2452 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2453                         hammer_tid_t delete_tid, uint32_t delete_ts,
2454                         int track, int64_t *stat_bytes)
2455 {
2456         struct hammer_btree_leaf_elm save_leaf;
2457         hammer_transaction_t trans;
2458         hammer_btree_leaf_elm_t leaf;
2459         hammer_node_t node;
2460         hammer_btree_elm_t elm;
2461         hammer_off_t data_offset;
2462         int32_t data_len;
2463         int64_t bytes;
2464         int ndelete;
2465         int error;
2466         int icount;
2467         int doprop;
2468
2469         error = hammer_cursor_upgrade(cursor);
2470         if (error)
2471                 return(error);
2472
2473         trans = cursor->trans;
2474         node = cursor->node;
2475         elm = &node->ondisk->elms[cursor->index];
2476         leaf = &elm->leaf;
2477         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2478
2479         hammer_sync_lock_sh(trans);
2480         bytes = 0;
2481         doprop = 0;
2482         icount = 0;
2483
2484         /*
2485          * Adjust the delete_tid.  Update the mirror_tid propagation field
2486          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2487          */
2488         if (delete_flags & HAMMER_DELETE_ADJUST) {
2489                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2490                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2491                                 icount = -1;
2492                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2493                                 icount = 1;
2494                 }
2495
2496                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2497                 elm->leaf.base.delete_tid = delete_tid;
2498                 elm->leaf.delete_ts = delete_ts;
2499                 hammer_modify_node_done(node);
2500
2501                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2502                         hammer_modify_node_field(trans, node, mirror_tid);
2503                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2504                         hammer_modify_node_done(node);
2505                         doprop = 1;
2506                         if (hammer_debug_general & 0x0002) {
2507                                 hdkprintf("propagate %016jx @%016jx\n",
2508                                         (intmax_t)elm->leaf.base.delete_tid,
2509                                         (intmax_t)node->node_offset);
2510                         }
2511                 }
2512
2513                 /*
2514                  * Adjust for the iteration.  We have deleted the current
2515                  * element and want to clear ATEDISK so the iteration does
2516                  * not skip the element after, which now becomes the current
2517                  * element.  This element must be re-tested if doing an
2518                  * iteration, which is handled by the RETEST flag.
2519                  */
2520                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2521                         cursor->flags |= HAMMER_CURSOR_RETEST;
2522                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2523                 }
2524
2525                 /*
2526                  * An on-disk record cannot have the same delete_tid
2527                  * as its create_tid.  In a chain of record updates
2528                  * this could result in a duplicate record.
2529                  */
2530                 KKASSERT(elm->leaf.base.delete_tid !=
2531                          elm->leaf.base.create_tid);
2532         }
2533
2534         /*
2535          * Destroy the B-Tree element if asked (typically if a nohistory
2536          * file or mount, or when called by the pruning code).
2537          *
2538          * Adjust the ATEDISK flag to properly support iterations.
2539          */
2540         if (delete_flags & HAMMER_DELETE_DESTROY) {
2541                 data_offset = elm->leaf.data_offset;
2542                 data_len = elm->leaf.data_len;
2543                 if (doprop) {
2544                         save_leaf = elm->leaf;
2545                         leaf = &save_leaf;
2546                 }
2547                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2548                     elm->leaf.base.delete_tid == 0) {
2549                         icount = -1;
2550                 }
2551
2552                 error = hammer_btree_delete(cursor, &ndelete);
2553                 if (error == 0) {
2554                         /*
2555                          * The deletion moves the next element (if any) to
2556                          * the current element position.  We must clear
2557                          * ATEDISK so this element is not skipped and we
2558                          * must set RETEST to force any iteration to re-test
2559                          * the element.
2560                          */
2561                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2562                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2563                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2564                         }
2565                         bytes += (ndelete * sizeof(struct hammer_node_ondisk));
2566
2567                         switch(HAMMER_ZONE(data_offset)) {
2568                         case HAMMER_ZONE_LARGE_DATA:
2569                         case HAMMER_ZONE_SMALL_DATA:
2570                         case HAMMER_ZONE_META:
2571                                 hammer_blockmap_free(trans,
2572                                                      data_offset, data_len);
2573                                 bytes += data_len;
2574                                 break;
2575                         default:
2576                                 break;
2577                         }
2578                 }
2579         }
2580
2581         /*
2582          * Track inode count and next_tid.  This is used by the mirroring
2583          * and PFS code.  icount can be negative, zero, or positive.
2584          */
2585         if (error == 0 && track) {
2586                 if (icount) {
2587                         hammer_modify_volume_field(trans, trans->rootvol,
2588                                                    vol0_stat_inodes);
2589                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2590                         hammer_modify_volume_done(trans->rootvol);
2591                 }
2592                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2593                         hammer_modify_volume_noundo(trans, trans->rootvol);
2594                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2595                         hammer_modify_volume_done(trans->rootvol);
2596                 }
2597         }
2598
2599         /*
2600          * mirror_tid propagation occurs if the node's mirror_tid had to be
2601          * updated while adjusting the delete_tid.
2602          *
2603          * This occurs when deleting even in nohistory mode, but does not
2604          * occur when pruning an already-deleted node.
2605          *
2606          * cursor->ip is NULL when called from the pruning, mirroring,
2607          * and pfs code.  If non-NULL propagation will be conditionalized
2608          * on whether the PFS is in no-history mode or not.
2609          *
2610          * WARNING: cursor's leaf pointer may have changed after do_propagation
2611          *          returns!
2612          */
2613         if (doprop) {
2614                 if (cursor->ip)
2615                         hammer_btree_do_propagation(cursor, leaf);
2616                 else
2617                         hammer_btree_do_propagation(cursor, leaf);
2618         }
2619         if (stat_bytes)
2620                 *stat_bytes = bytes;
2621         hammer_sync_unlock(trans);
2622         return (error);
2623 }
2624
2625 /*
2626  * Determine whether we can remove a directory.  This routine checks whether
2627  * a directory is empty or not and enforces flush connectivity.
2628  *
2629  * Flush connectivity requires that we block if the target directory is
2630  * currently flushing, otherwise it may not end up in the same flush group.
2631  *
2632  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2633  */
2634 int
2635 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2636 {
2637         struct hammer_cursor cursor;
2638         int error;
2639
2640         /*
2641          * Check directory empty
2642          */
2643         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2644
2645         cursor.key_beg.localization = ip->obj_localization |
2646                                       hammer_dir_localization(ip);
2647         cursor.key_beg.obj_id = ip->obj_id;
2648         cursor.key_beg.create_tid = 0;
2649         cursor.key_beg.delete_tid = 0;
2650         cursor.key_beg.obj_type = 0;
2651         cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START;
2652         cursor.key_beg.key = HAMMER_MIN_KEY;
2653
2654         cursor.key_end = cursor.key_beg;
2655         cursor.key_end.rec_type = HAMMER_RECTYPE_MAX;
2656         cursor.key_end.key = HAMMER_MAX_KEY;
2657
2658         cursor.asof = ip->obj_asof;
2659         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2660
2661         error = hammer_ip_first(&cursor);
2662         if (error == ENOENT)
2663                 error = 0;
2664         else if (error == 0)
2665                 error = ENOTEMPTY;
2666         hammer_done_cursor(&cursor);
2667         return(error);
2668 }
2669
2670 /*
2671  * Localize the data payload.  Directory entries may need their
2672  * localization adjusted.
2673  */
2674 static
2675 int
2676 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2677                             hammer_btree_leaf_elm_t leaf)
2678 {
2679         uint32_t localization;
2680
2681         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2682                 localization = leaf->base.localization &
2683                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2684                 if (data->entry.localization != localization) {
2685                         data->entry.localization = localization;
2686                         hammer_crc_set_leaf(data, leaf);
2687                 }
2688         }
2689         return(0);
2690 }