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(struct hammer_record *record)
 365 {
 366         hammer_mount_t hmp;
 367         hammer_reserve_t resv;
 368         hammer_inode_t ip;
 369         hammer_inode_t target_ip;
 370         int diddrop;
 371
 372         hammer_rel(&record->lock);
 373
 374         if (hammer_norefs(&record->lock)) {
 375                 /*
 376                  * Upon release of the last reference wakeup any waiters.
 377                  * The record structure may get destroyed so callers will
 378                  * loop up and do a relookup.
 379                  *
 380                  * WARNING!  Record must be removed from RB-TREE before we
 381                  * might possibly block.  hammer_test_inode() can block!
 382                  */
 383                 ip = record->ip;
 384                 hmp = ip->hmp;
 385
 386                 /*
 387                  * Upon release of the last reference a record marked deleted
 388                  * by the front or backend, or committed by the backend,
 389                  * is destroyed.
 390                  */
 391                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 392                                      HAMMER_RECF_DELETED_BE |
 393                                      HAMMER_RECF_COMMITTED)) {
 394                         KKASSERT(hammer_isactive(&ip->lock) > 0);
 395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 396
 397                         /*
 398                          * target_ip may have zero refs, we have to ref it
 399                          * to prevent it from being ripped out from under
 400                          * us.
 401                          */
 402                         if ((target_ip = record->target_ip) != NULL) {
 403                                 TAILQ_REMOVE(&target_ip->target_list,
 404                                              record, target_entry);
 405                                 record->target_ip = NULL;
 406                                 hammer_ref(&target_ip->lock);
 407                         }
 408
 409                         /*
 410                          * Remove the record from the 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(struct hammer_transaction *trans,
 660                      struct hammer_inode *dip, const char *name, int bytes,
 661                      struct hammer_inode *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                         error = ENOSPC;
 714                         goto failed;
 715                 }
 716         }
 717
 718         /*
 719          * The target inode and the directory entry are bound together.
 720          */
 721         record->target_ip = ip;
 722         record->flush_state = HAMMER_FST_SETUP;
 723         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 724
 725         /*
 726          * The inode now has a dependancy and must be taken out of the idle
 727          * state.  An inode not in an idle state is given an extra reference.
 728          *
 729          * When transitioning to a SETUP state flag for an automatic reflush
 730          * when the dependancies are disposed of if someone is waiting on
 731          * the inode.
 732          */
 733         if (ip->flush_state == HAMMER_FST_IDLE) {
 734                 hammer_ref(&ip->lock);
 735                 ip->flush_state = HAMMER_FST_SETUP;
 736                 if (ip->flags & HAMMER_INODE_FLUSHW)
 737                         ip->flags |= HAMMER_INODE_REFLUSH;
 738         }
 739         error = hammer_mem_add(record);
 740         if (error == 0) {
 741                 dip->ino_data.mtime = trans->time;
 742                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 743         }
 744 failed:
 745         hammer_done_cursor(&cursor);
 746         return(error);
 747 }
 748
 749 /*
 750  * Delete the directory entry and update the inode link count.  The
 751  * cursor must be seeked to the directory entry record being deleted.
 752  *
 753  * The related inode should be share-locked by the caller.  The caller is
 754  * on the frontend.  It could also be NULL indicating that the directory
 755  * entry being removed has no related inode.
 756  *
 757  * This function can return EDEADLK requiring the caller to terminate
 758  * the cursor, any locks, wait on the returned record, and retry.
 759  */
 760 int
 761 hammer_ip_del_direntry(struct hammer_transaction *trans,
 762                      hammer_cursor_t cursor, struct hammer_inode *dip,
 763                      struct hammer_inode *ip)
 764 {
 765         hammer_record_t record;
 766         int error;
 767
 768         if (hammer_cursor_inmem(cursor)) {
 769                 /*
 770                  * In-memory (unsynchronized) records can simply be freed.
 771                  *
 772                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 773                  * by the backend, we must still avoid races against the
 774                  * backend potentially syncing the record to the media.
 775                  *
 776                  * We cannot call hammer_ip_delete_record(), that routine may
 777                  * only be called from the backend.
 778                  */
 779                 record = cursor->iprec;
 780                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 781                                      HAMMER_RECF_DELETED_BE |
 782                                      HAMMER_RECF_COMMITTED)) {
 783                         KKASSERT(cursor->deadlk_rec == NULL);
 784                         hammer_ref(&record->lock);
 785                         cursor->deadlk_rec = record;
 786                         error = EDEADLK;
 787                 } else {
 788                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 789                         record->flags |= HAMMER_RECF_DELETED_FE;
 790                         error = 0;
 791                 }
 792         } else {
 793                 /*
 794                  * If the record is on-disk we have to queue the deletion by
 795                  * the record's key.  This also causes lookups to skip the
 796                  * record (lookups for the purposes of finding an unused
 797                  * directory key do not skip the record).
 798                  */
 799                 KKASSERT(dip->flags &
 800                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 801                 record = hammer_alloc_mem_record(dip, 0);
 802                 record->type = HAMMER_MEM_RECORD_DEL;
 803                 record->leaf.base = cursor->leaf->base;
 804                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
 805
 806                 /*
 807                  * ip may be NULL, indicating the deletion of a directory
 808                  * entry which has no related inode.
 809                  */
 810                 record->target_ip = ip;
 811                 if (ip) {
 812                         record->flush_state = HAMMER_FST_SETUP;
 813                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 814                                           target_entry);
 815                 } else {
 816                         record->flush_state = HAMMER_FST_IDLE;
 817                 }
 818
 819                 /*
 820                  * The inode now has a dependancy and must be taken out of
 821                  * the idle state.  An inode not in an idle state is given
 822                  * an extra reference.
 823                  *
 824                  * When transitioning to a SETUP state flag for an automatic
 825                  * reflush when the dependancies are disposed of if someone
 826                  * is waiting on the inode.
 827                  */
 828                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 829                         hammer_ref(&ip->lock);
 830                         ip->flush_state = HAMMER_FST_SETUP;
 831                         if (ip->flags & HAMMER_INODE_FLUSHW)
 832                                 ip->flags |= HAMMER_INODE_REFLUSH;
 833                 }
 834
 835                 error = hammer_mem_add(record);
 836         }
 837
 838         /*
 839          * One less link.  The file may still be open in the OS even after
 840          * all links have gone away.
 841          *
 842          * We have to terminate the cursor before syncing the inode to
 843          * avoid deadlocking against ourselves.  XXX this may no longer
 844          * be true.
 845          *
 846          * If nlinks drops to zero and the vnode is inactive (or there is
 847          * no vnode), call hammer_inode_unloadable_check() to zonk the
 848          * inode.  If we don't do this here the inode will not be destroyed
 849          * on-media until we unmount.
 850          */
 851         if (error == 0) {
 852                 if (ip) {
 853                         --ip->ino_data.nlinks;  /* do before we might block */
 854                         ip->ino_data.ctime = trans->time;
 855                 }
 856                 dip->ino_data.mtime = trans->time;
 857                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 858                 if (ip) {
 859                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 860                         if (ip->ino_data.nlinks == 0 &&
 861                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 862                                 hammer_done_cursor(cursor);
 863                                 hammer_inode_unloadable_check(ip, 1);
 864                                 hammer_flush_inode(ip, 0);
 865                         }
 866                 }
 867
 868         }
 869         return(error);
 870 }
 871
 872 /*
 873  * Add a record to an inode.
 874  *
 875  * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and
 876  * initialize the following additional fields that are not initialized by these
 877  * functions.
 878  *
 879  * The related inode should be share-locked by the caller.  The caller is
 880  * on the frontend.
 881  *
 882  * record->leaf.base.key
 883  * record->leaf.base.rec_type
 884  * record->leaf.base.localization
 885  */
 886 int
 887 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 888 {
 889         hammer_inode_t ip = record->ip;
 890         int error;
 891
 892         KKASSERT(record->leaf.base.localization != 0);
 893         record->leaf.base.obj_id = ip->obj_id;
 894         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 895         error = hammer_mem_add(record);
 896         return(error);
 897 }
 898
 899 /*
 900  * Locate a pre-existing bulk record in memory.  The caller wishes to
 901  * replace the record with a new one.  The existing record may have a
 902  * different length (and thus a different key) so we have to use an
 903  * overlap check function.
 904  */
 905 static hammer_record_t
 906 hammer_ip_get_bulk(hammer_record_t record)
 907 {
 908         struct hammer_bulk_info info;
 909         hammer_inode_t ip = record->ip;
 910
 911         info.record = record;
 912         info.conflict = NULL;
 913         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 914                                    hammer_bulk_scan_callback, &info);
 915
 916         return(info.conflict);  /* may be NULL */
 917 }
 918
 919 /*
 920  * Take records vetted by overlap_cmp.  The first non-deleted record
 921  * (if any) stops the scan.
 922  */
 923 static int
 924 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 925 {
 926         struct hammer_bulk_info *info = data;
 927
 928         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 929                              HAMMER_RECF_COMMITTED)) {
 930                 return(0);
 931         }
 932         hammer_ref(&record->lock);
 933         info->conflict = record;
 934         return(-1);                     /* stop scan */
 935 }
 936
 937 /*
 938  * Reserve blockmap space placemarked with an in-memory record.
 939  *
 940  * This routine is called by the frontend in order to be able to directly
 941  * flush a buffer cache buffer.  The frontend has locked the related buffer
 942  * cache buffers and we should be able to manipulate any overlapping
 943  * in-memory records.
 944  *
 945  * The caller is responsible for adding the returned record and deleting
 946  * the returned conflicting record (if any), typically by calling
 947  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
 948  */
 949 hammer_record_t
 950 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 951                    int *errorp)
 952 {
 953         hammer_record_t record;
 954         hammer_dedup_cache_t dcp;
 955         hammer_crc_t crc;
 956         int zone;
 957
 958         /*
 959          * Create a record to cover the direct write.  The record cannot
 960          * be added to the in-memory RB tree here as it might conflict
 961          * with an existing memory record.  See hammer_io_direct_write().
 962          *
 963          * The backend is responsible for finalizing the space reserved in
 964          * this record.
 965          *
 966          * XXX bytes not aligned, depend on the reservation code to
 967          * align the reservation.
 968          */
 969         record = hammer_alloc_mem_record(ip, 0);
 970         zone = hammer_data_zone_index(bytes);
 971         if (bytes == 0)
 972                 crc = 0;
 973         else
 974                 crc = crc32(data, bytes);
 975
 976         if (hammer_live_dedup == 0)
 977                 goto nodedup;
 978         if ((dcp = hammer_dedup_cache_lookup(ip->hmp, crc)) != NULL) {
 979                 struct hammer_dedup_cache tmp = *dcp;
 980
 981                 record->resv = hammer_blockmap_reserve_dedup(ip->hmp, zone,
 982                         bytes, tmp.data_offset, errorp);
 983                 if (record->resv == NULL)
 984                         goto nodedup;
 985
 986                 if (!hammer_dedup_validate(&tmp, zone, bytes, data)) {
 987                         hammer_blockmap_reserve_complete(ip->hmp, record->resv);
 988                         goto nodedup;
 989                 }
 990
 991                 record->leaf.data_offset = tmp.data_offset;
 992                 record->flags |= HAMMER_RECF_DEDUPED;
 993         } else {
 994 nodedup:
 995                 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 996                        &record->leaf.data_offset, errorp);
 997                 if (record->resv == NULL) {
 998                         hdkprintf("reservation failed\n");
 999                         hammer_rel_mem_record(record);
1000                         return(NULL);
1001                 }
1002         }
1003
1004         record->type = HAMMER_MEM_RECORD_DATA;
1005         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
1006         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
1007         record->leaf.base.obj_id = ip->obj_id;
1008         record->leaf.base.key = file_offset + bytes;
1009         record->leaf.base.localization = ip->obj_localization |
1010                                          HAMMER_LOCALIZE_MISC;
1011         record->leaf.data_len = bytes;
1012         record->leaf.data_crc = crc;
1013         KKASSERT(*errorp == 0);
1014
1015         return(record);
1016 }
1017
1018 /*
1019  * Called by hammer_io_direct_write() prior to any possible completion
1020  * of the BIO to emplace the memory record associated with the I/O and
1021  * to replace any prior memory record which might still be active.
1022  *
1023  * Setting the FE deleted flag on the old record (if any) avoids any RB
1024  * tree insertion conflict, amoung other things.
1025  *
1026  * This has to be done prior to the caller completing any related buffer
1027  * cache I/O or a reinstantiation of the buffer may load data from the
1028  * old media location instead of the new media location.  The holding
1029  * of the locked buffer cache buffer serves to interlock the record
1030  * replacement operation.
1031  */
1032 void
1033 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1034 {
1035         hammer_record_t conflict;
1036         int error __debugvar;
1037
1038         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1039                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1040                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1041                         break;
1042                 }
1043                 conflict->flags |= HAMMER_RECF_WANTED;
1044                 tsleep(conflict, 0, "hmrrc3", 0);
1045                 hammer_rel_mem_record(conflict);
1046         }
1047         error = hammer_mem_add(record);
1048         if (conflict)
1049                 hammer_rel_mem_record(conflict);
1050         KKASSERT(error == 0);
1051 }
1052
1053 /*
1054  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1055  * and records in a flushing state are handled by the backend.  The vnops
1056  * setattr code will handle the block containing the truncation point.
1057  *
1058  * Partial blocks are not deleted.
1059  *
1060  * This code is only called on regular files.
1061  */
1062 int
1063 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1064 {
1065         struct rec_trunc_info info;
1066
1067         switch(ip->ino_data.obj_type) {
1068         case HAMMER_OBJTYPE_REGFILE:
1069                 info.rec_type = HAMMER_RECTYPE_DATA;
1070                 break;
1071         case HAMMER_OBJTYPE_DBFILE:
1072                 info.rec_type = HAMMER_RECTYPE_DB;
1073                 break;
1074         default:
1075                 return(EINVAL);
1076         }
1077         info.trunc_off = file_size;
1078         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1079                                    hammer_frontend_trunc_callback, &info);
1080         return(0);
1081 }
1082
1083 /*
1084  * Scan callback for frontend records to destroy during a truncation.
1085  * We must ensure that DELETED_FE is set on the record or the frontend
1086  * will get confused in future read() calls.
1087  *
1088  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1089  *       In this rare case we must wait for the interlock to be cleared.
1090  *
1091  * NOTE: This function is only called on regular files.  There are further
1092  *       restrictions to the setting of DELETED_FE on directory records
1093  *       undergoing a flush due to sensitive inode link count calculations.
1094  */
1095 static int
1096 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1097 {
1098         if (record->flags & HAMMER_RECF_DELETED_FE)
1099                 return(0);
1100 #if 0
1101         if (record->flush_state == HAMMER_FST_FLUSH)
1102                 return(0);
1103 #endif
1104         hammer_ref(&record->lock);
1105         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1106                 hammer_wait_mem_record_ident(record, "hmmtrr");
1107         record->flags |= HAMMER_RECF_DELETED_FE;
1108         hammer_rel_mem_record(record);
1109         return(0);
1110 }
1111
1112 /*
1113  * Return 1 if the caller must check for and delete existing records
1114  * before writing out a new data record.
1115  *
1116  * Return 0 if the caller can just insert the record into the B-Tree without
1117  * checking.
1118  */
1119 static int
1120 hammer_record_needs_overwrite_delete(hammer_record_t record)
1121 {
1122         hammer_inode_t ip = record->ip;
1123         int64_t file_offset;
1124         int r;
1125
1126         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1127                 file_offset = record->leaf.base.key;
1128         else
1129                 file_offset = record->leaf.base.key - record->leaf.data_len;
1130         r = (file_offset < ip->save_trunc_off);
1131         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1132                 if (ip->save_trunc_off <= record->leaf.base.key)
1133                         ip->save_trunc_off = record->leaf.base.key + 1;
1134         } else {
1135                 if (ip->save_trunc_off < record->leaf.base.key)
1136                         ip->save_trunc_off = record->leaf.base.key;
1137         }
1138         return(r);
1139 }
1140
1141 /*
1142  * Backend code.  Sync a record to the media.
1143  */
1144 int
1145 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1146 {
1147         hammer_transaction_t trans = cursor->trans;
1148         int64_t file_offset;
1149         int bytes;
1150         void *bdata;
1151         int error;
1152         int doprop;
1153
1154         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1155         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1156         KKASSERT(record->leaf.base.localization != 0);
1157
1158         /*
1159          * Any direct-write related to the record must complete before we
1160          * can sync the record to the on-disk media.
1161          */
1162         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1163                 hammer_io_direct_wait(record);
1164
1165         /*
1166          * If this is a bulk-data record placemarker there may be an existing
1167          * record on-disk, indicating a data overwrite.  If there is the
1168          * on-disk record must be deleted before we can insert our new record.
1169          *
1170          * We've synthesized this record and do not know what the create_tid
1171          * on-disk is, nor how much data it represents.
1172          *
1173          * Keep in mind that (key) for data records is (base_offset + len),
1174          * not (base_offset).  Also, we only want to get rid of on-disk
1175          * records since we are trying to sync our in-memory record, call
1176          * hammer_ip_delete_range() with truncating set to 1 to make sure
1177          * it skips in-memory records.
1178          *
1179          * It is ok for the lookup to return ENOENT.
1180          *
1181          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1182          * to call hammer_ip_delete_range() or not.  This also means we must
1183          * update sync_trunc_off() as we write.
1184          */
1185         if (record->type == HAMMER_MEM_RECORD_DATA &&
1186             hammer_record_needs_overwrite_delete(record)) {
1187                 file_offset = record->leaf.base.key - record->leaf.data_len;
1188                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1189                         ~HAMMER_BUFMASK;
1190                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1191                 error = hammer_ip_delete_range(
1192                                 cursor, record->ip,
1193                                 file_offset, file_offset + bytes - 1,
1194                                 1);
1195                 if (error && error != ENOENT)
1196                         goto done;
1197         }
1198
1199         /*
1200          * If this is a general record there may be an on-disk version
1201          * that must be deleted before we can insert the new record.
1202          */
1203         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1204                 error = hammer_delete_general(cursor, record->ip, &record->leaf);
1205                 if (error && error != ENOENT)
1206                         goto done;
1207         }
1208
1209         /*
1210          * Setup the cursor.
1211          */
1212         hammer_normalize_cursor(cursor);
1213         cursor->key_beg = record->leaf.base;
1214         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1215         cursor->flags |= HAMMER_CURSOR_BACKEND;
1216         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1217
1218         /*
1219          * Records can wind up on-media before the inode itself is on-media.
1220          * Flag the case.
1221          */
1222         record->ip->flags |= HAMMER_INODE_DONDISK;
1223
1224         /*
1225          * If we are deleting a directory entry an exact match must be
1226          * found on-disk.
1227          */
1228         if (record->type == HAMMER_MEM_RECORD_DEL) {
1229                 error = hammer_btree_lookup(cursor);
1230                 if (error == 0) {
1231                         KKASSERT(cursor->iprec == NULL);
1232                         error = hammer_ip_delete_record(cursor, record->ip,
1233                                                         trans->tid);
1234                         if (error == 0) {
1235                                 record->flags |= HAMMER_RECF_DELETED_BE |
1236                                                  HAMMER_RECF_COMMITTED;
1237                                 ++record->ip->rec_generation;
1238                         }
1239                 }
1240                 goto done;
1241         }
1242
1243         /*
1244          * We are inserting.
1245          *
1246          * Issue a lookup to position the cursor and locate the insertion
1247          * point.  The target key should not exist.  If we are creating a
1248          * directory entry we may have to iterate the low 32 bits of the
1249          * key to find an unused key.
1250          */
1251         hammer_sync_lock_sh(trans);
1252         cursor->flags |= HAMMER_CURSOR_INSERT;
1253         error = hammer_btree_lookup(cursor);
1254         if (hammer_debug_inode)
1255                 hdkprintf("DOINSERT LOOKUP %d\n", error);
1256         if (error == 0) {
1257                 hdkprintf("duplicate rec at (%016llx)\n",
1258                         (long long)record->leaf.base.key);
1259                 if (hammer_debug_critical)
1260                         Debugger("duplicate record1");
1261                 error = EIO;
1262         }
1263
1264         if (error != ENOENT)
1265                 goto done_unlock;
1266
1267         /*
1268          * Allocate the record and data.  The result buffers will be
1269          * marked as being modified and further calls to
1270          * hammer_modify_buffer() will result in unneeded UNDO records.
1271          *
1272          * Support zero-fill records (data == NULL and data_len != 0)
1273          */
1274         if (record->type == HAMMER_MEM_RECORD_DATA) {
1275                 /*
1276                  * The data portion of a bulk-data record has already been
1277                  * committed to disk, we need only adjust the layer2
1278                  * statistics in the same transaction as our B-Tree insert.
1279                  */
1280                 KKASSERT(record->leaf.data_offset != 0);
1281                 error = hammer_blockmap_finalize(trans,
1282                                                  record->resv,
1283                                                  record->leaf.data_offset,
1284                                                  record->leaf.data_len);
1285
1286                 if (hammer_live_dedup == 2 &&
1287                     (record->flags & HAMMER_RECF_DEDUPED) == 0) {
1288                         hammer_dedup_cache_add(record->ip, &record->leaf);
1289                 }
1290         } else if (record->data && record->leaf.data_len) {
1291                 /*
1292                  * Wholely cached record, with data.  Allocate the data.
1293                  */
1294                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1295                                           record->leaf.base.rec_type,
1296                                           &record->leaf.data_offset,
1297                                           &cursor->data_buffer,
1298                                           0, &error);
1299                 if (bdata == NULL)
1300                         goto done_unlock;
1301                 hammer_crc_set_leaf(record->data, &record->leaf);
1302                 hammer_modify_buffer_noundo(trans, cursor->data_buffer);
1303                 bcopy(record->data, bdata, record->leaf.data_len);
1304                 hammer_modify_buffer_done(cursor->data_buffer);
1305         } else {
1306                 /*
1307                  * Wholely cached record, without data.
1308                  */
1309                 record->leaf.data_offset = 0;
1310                 record->leaf.data_crc = 0;
1311         }
1312
1313         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1314         if (hammer_debug_inode && error) {
1315                 hdkprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1316                         error,
1317                         (long long)cursor->node->node_offset,
1318                         cursor->index,
1319                         (long long)record->leaf.base.key);
1320         }
1321
1322         /*
1323          * Our record is on-disk and we normally mark the in-memory version
1324          * as having been committed (and not BE-deleted).
1325          *
1326          * If the record represented a directory deletion but we had to
1327          * sync a valid directory entry to disk due to dependancies,
1328          * we must convert the record to a covering delete so the
1329          * frontend does not have visibility on the synced entry.
1330          *
1331          * WARNING: cursor's leaf pointer may have changed after do_propagation
1332          *          returns!
1333          */
1334         if (error == 0) {
1335                 if (doprop) {
1336                         hammer_btree_do_propagation(cursor,
1337                                                     record->ip->pfsm,
1338                                                     &record->leaf);
1339                 }
1340                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1341                         /*
1342                          * Must convert deleted directory entry add
1343                          * to a directory entry delete.
1344                          */
1345                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1346                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1347                         record->type = HAMMER_MEM_RECORD_DEL;
1348                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1349                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1350                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1351                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1352                                                  HAMMER_RECF_DELETED_BE)) == 0);
1353                         /* converted record is not yet committed */
1354                         /* hammer_flush_record_done takes care of the rest */
1355                 } else {
1356                         /*
1357                          * Everything went fine and we are now done with
1358                          * this record.
1359                          */
1360                         record->flags |= HAMMER_RECF_COMMITTED;
1361                         ++record->ip->rec_generation;
1362                 }
1363         } else {
1364                 if (record->leaf.data_offset) {
1365                         hammer_blockmap_free(trans, record->leaf.data_offset,
1366                                              record->leaf.data_len);
1367                 }
1368         }
1369 done_unlock:
1370         hammer_sync_unlock(trans);
1371 done:
1372         return(error);
1373 }
1374
1375 /*
1376  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1377  * entry's key is used to deal with hash collisions in the upper 32 bits.
1378  * A unique 64 bit key is generated in-memory and may be regenerated a
1379  * second time when the directory record is flushed to the on-disk B-Tree.
1380  *
1381  * A referenced record is passed to this function.  This function
1382  * eats the reference.  If an error occurs the record will be deleted.
1383  *
1384  * A copy of the temporary record->data pointer provided by the caller
1385  * will be made.
1386  */
1387 int
1388 hammer_mem_add(hammer_record_t record)
1389 {
1390         hammer_mount_t hmp = record->ip->hmp;
1391
1392         /*
1393          * Make a private copy of record->data
1394          */
1395         if (record->data)
1396                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1397
1398         /*
1399          * Insert into the RB tree.  A unique key should have already
1400          * been selected if this is a directory entry.
1401          */
1402         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1403                 record->flags |= HAMMER_RECF_DELETED_FE;
1404                 hammer_rel_mem_record(record);
1405                 return (EEXIST);
1406         }
1407         ++hmp->rsv_recs;
1408         ++record->ip->rsv_recs;
1409         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1410         record->flags |= HAMMER_RECF_ONRBTREE;
1411         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1412         hammer_rel_mem_record(record);
1413         return(0);
1414 }
1415
1416 /************************************************************************
1417  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1418  ************************************************************************
1419  *
1420  * These functions augment the B-Tree scanning functions in hammer_btree.c
1421  * by merging in-memory records with on-disk records.
1422  */
1423
1424 /*
1425  * Locate a particular record either in-memory or on-disk.
1426  *
1427  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1428  * NOT be called to iterate results.
1429  */
1430 int
1431 hammer_ip_lookup(hammer_cursor_t cursor)
1432 {
1433         int error;
1434
1435         /*
1436          * If the element is in-memory return it without searching the
1437          * on-disk B-Tree
1438          */
1439         KKASSERT(cursor->ip);
1440         error = hammer_mem_lookup(cursor);
1441         if (error == 0) {
1442                 cursor->leaf = &cursor->iprec->leaf;
1443                 return(error);
1444         }
1445         if (error != ENOENT)
1446                 return(error);
1447
1448         /*
1449          * If the inode has on-disk components search the on-disk B-Tree.
1450          */
1451         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1452                 return(error);
1453         error = hammer_btree_lookup(cursor);
1454         if (error == 0)
1455                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1456         return(error);
1457 }
1458
1459 /*
1460  * Helper for hammer_ip_first()/hammer_ip_next()
1461  *
1462  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1463  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1464  * state if hammer_ip_next() needs to re-seek.
1465  */
1466 static __inline
1467 int
1468 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1469 {
1470         hammer_inode_t ip = cursor->ip;
1471         int error;
1472
1473         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1474                 error = hammer_btree_lookup(cursor);
1475                 if (error == ENOENT || error == EDEADLK) {
1476                         if (hammer_debug_general & 0x2000) {
1477                                 hdkprintf("error %d node %p %016llx index %d\n",
1478                                         error, cursor->node,
1479                                         (long long)cursor->node->node_offset,
1480                                         cursor->index);
1481                         }
1482                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1483                         error = hammer_btree_iterate(cursor);
1484                 }
1485                 if (error == 0) {
1486                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1487                                            HAMMER_CURSOR_ATEDISK);
1488                 } else {
1489                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1490                                          HAMMER_CURSOR_ATEDISK;
1491                         if (error == ENOENT)
1492                                 error = 0;
1493                 }
1494         } else {
1495                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1496                 error = 0;
1497         }
1498         return(error);
1499 }
1500
1501 /*
1502  * Helper for hammer_ip_next()
1503  *
1504  * The caller has determined that the media cursor is further along than the
1505  * memory cursor and must be reseeked after a generation number change.
1506  */
1507 static
1508 int
1509 _hammer_ip_reseek(hammer_cursor_t cursor)
1510 {
1511         struct hammer_base_elm save;
1512         hammer_btree_elm_t elm;
1513         int error __debugvar;
1514         int r;
1515         int again = 0;
1516
1517         /*
1518          * Do the re-seek.
1519          */
1520         hkprintf("Debug: re-seeked during scan @ino=%016llx\n",
1521                 (long long)cursor->ip->obj_id);
1522         save = cursor->key_beg;
1523         cursor->key_beg = cursor->iprec->leaf.base;
1524         error = _hammer_ip_seek_btree(cursor);
1525         KKASSERT(error == 0);
1526         cursor->key_beg = save;
1527
1528         /*
1529          * If the memory record was previous returned to
1530          * the caller and the media record matches
1531          * (-1/+1: only create_tid differs), then iterate
1532          * the media record to avoid a double result.
1533          */
1534         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1535             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1536                 elm = &cursor->node->ondisk->elms[cursor->index];
1537                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1538                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1539                         if (r >= -1 && r <= 1) {
1540                                 hkprintf("Debug: iterated after "
1541                                         "re-seek (asof r=%d)\n", r);
1542                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1543                                 again = 1;
1544                         }
1545                 } else {
1546                         if (r == 0) {
1547                                 hkprintf("Debug: iterated after "
1548                                         "re-seek\n");
1549                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1550                                 again = 1;
1551                         }
1552                 }
1553         }
1554         return(again);
1555 }
1556
1557 /*
1558  * Locate the first record within the cursor's key_beg/key_end range,
1559  * restricted to a particular inode.  0 is returned on success, ENOENT
1560  * if no records matched the requested range, or some other error.
1561  *
1562  * When 0 is returned hammer_ip_next() may be used to iterate additional
1563  * records within the requested range.
1564  *
1565  * This function can return EDEADLK, requiring the caller to terminate
1566  * the cursor and try again.
1567  */
1568
1569 int
1570 hammer_ip_first(hammer_cursor_t cursor)
1571 {
1572         hammer_inode_t ip __debugvar = cursor->ip;
1573         int error;
1574
1575         KKASSERT(ip != NULL);
1576
1577         /*
1578          * Clean up fields and setup for merged scan
1579          */
1580         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1581
1582         /*
1583          * Search the in-memory record list (Red-Black tree).  Unlike the
1584          * B-Tree search, mem_first checks for records in the range.
1585          *
1586          * This function will setup both ATEMEM and MEMEOF properly for
1587          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1588          */
1589         hammer_mem_first(cursor);
1590
1591         /*
1592          * Detect generation changes during blockages, including
1593          * blockages which occur on the initial btree search.
1594          */
1595         cursor->rec_generation = cursor->ip->rec_generation;
1596
1597         /*
1598          * Initial search and result
1599          */
1600         error = _hammer_ip_seek_btree(cursor);
1601         if (error == 0)
1602                 error = hammer_ip_next(cursor);
1603
1604         return (error);
1605 }
1606
1607 /*
1608  * Retrieve the next record in a merged iteration within the bounds of the
1609  * cursor.  This call may be made multiple times after the cursor has been
1610  * initially searched with hammer_ip_first().
1611  *
1612  * There are numerous special cases in this code to deal with races between
1613  * in-memory records and on-media records.
1614  *
1615  * 0 is returned on success, ENOENT if no further records match the
1616  * requested range, or some other error code is returned.
1617  */
1618 int
1619 hammer_ip_next(hammer_cursor_t cursor)
1620 {
1621         hammer_btree_elm_t elm;
1622         hammer_record_t rec;
1623         hammer_record_t tmprec;
1624         int error;
1625         int r;
1626
1627 again:
1628         /*
1629          * Get the next on-disk record
1630          *
1631          * NOTE: If we deleted the last on-disk record we had scanned
1632          *       ATEDISK will be clear and RETEST will be set, forcing
1633          *       a call to iterate.  The fact that ATEDISK is clear causes
1634          *       iterate to re-test the 'current' element.  If ATEDISK is
1635          *       set, iterate will skip the 'current' element.
1636          */
1637         error = 0;
1638         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1639                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1640                                      HAMMER_CURSOR_RETEST)) {
1641                         error = hammer_btree_iterate(cursor);
1642                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1643                         if (error == 0) {
1644                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1645                                 hammer_cache_node(&cursor->ip->cache[1],
1646                                                   cursor->node);
1647                         } else if (error == ENOENT) {
1648                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1649                                                  HAMMER_CURSOR_ATEDISK;
1650                                 error = 0;
1651                         }
1652                 }
1653         }
1654
1655         /*
1656          * If the generation changed the backend has deleted or committed
1657          * one or more memory records since our last check.
1658          *
1659          * When this case occurs if the disk cursor is > current memory record
1660          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1661          * Since the cursor is ahead it must have not yet been eaten (if
1662          * not at eof anyway). (XXX data offset case?)
1663          *
1664          * NOTE: we are not doing a full check here.  That will be handled
1665          * later on.
1666          *
1667          * If we have exhausted all memory records we do not have to do any
1668          * further seeks.
1669          */
1670         while (cursor->rec_generation != cursor->ip->rec_generation &&
1671                error == 0) {
1672                 hkprintf("Debug: generation changed during scan @ino=%016llx\n",
1673                         (long long)cursor->ip->obj_id);
1674                 cursor->rec_generation = cursor->ip->rec_generation;
1675                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1676                         break;
1677                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1678                         r = 1;
1679                 } else {
1680                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1681                         elm = &cursor->node->ondisk->elms[cursor->index];
1682                         r = hammer_btree_cmp(&elm->base,
1683                                              &cursor->iprec->leaf.base);
1684                 }
1685
1686                 /*
1687                  * Do we re-seek the media cursor?
1688                  */
1689                 if (r > 0) {
1690                         if (_hammer_ip_reseek(cursor))
1691                                 goto again;
1692                 }
1693         }
1694
1695         /*
1696          * We can now safely get the next in-memory record.  We cannot
1697          * block here.
1698          *
1699          * hammer_rec_scan_cmp:  Is the record still in our general range,
1700          *                       (non-inclusive of snapshot exclusions)?
1701          * hammer_rec_scan_callback: Is the record in our snapshot?
1702          */
1703         tmprec = NULL;
1704         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1705                 /*
1706                  * If the current memory record was eaten then get the next
1707                  * one.  Stale records are skipped.
1708                  */
1709                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1710                         tmprec = cursor->iprec;
1711                         cursor->iprec = NULL;
1712                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1713                         while (rec) {
1714                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1715                                         break;
1716                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1717                                         break;
1718                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1719                         }
1720                         if (cursor->iprec) {
1721                                 KKASSERT(cursor->iprec == rec);
1722                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1723                         } else {
1724                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1725                         }
1726                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1727                 }
1728         }
1729
1730         /*
1731          * MEMORY RECORD VALIDITY TEST
1732          *
1733          * (We still can't block, which is why tmprec is being held so
1734          * long).
1735          *
1736          * If the memory record is no longer valid we skip it.  It may
1737          * have been deleted by the frontend.  If it was deleted or
1738          * committed by the backend the generation change re-seeked the
1739          * disk cursor and the record will be present there.
1740          */
1741         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1742                 KKASSERT(cursor->iprec);
1743                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1744                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1745                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1746                         if (tmprec)
1747                                 hammer_rel_mem_record(tmprec);
1748                         goto again;
1749                 }
1750         }
1751         if (tmprec)
1752                 hammer_rel_mem_record(tmprec);
1753
1754         /*
1755          * Extract either the disk or memory record depending on their
1756          * relative position.
1757          */
1758         error = 0;
1759         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1760         case 0:
1761                 /*
1762                  * Both entries valid.   Compare the entries and nominally
1763                  * return the first one in the sort order.  Numerous cases
1764                  * require special attention, however.
1765                  */
1766                 elm = &cursor->node->ondisk->elms[cursor->index];
1767                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1768
1769                 /*
1770                  * If the two entries differ only by their key (-2/2) or
1771                  * create_tid (-1/1), and are DATA records, we may have a
1772                  * nominal match.  We have to calculate the base file
1773                  * offset of the data.
1774                  */
1775                 if (r <= 2 && r >= -2 && r != 0 &&
1776                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1777                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1778                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1779                         int64_t base2 = cursor->iprec->leaf.base.key -
1780                                         cursor->iprec->leaf.data_len;
1781                         if (base1 == base2)
1782                                 r = 0;
1783                 }
1784
1785                 if (r < 0) {
1786                         error = hammer_btree_extract(cursor,
1787                                                      HAMMER_CURSOR_GET_LEAF);
1788                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1789                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1790                         break;
1791                 }
1792
1793                 /*
1794                  * If the entries match exactly the memory entry is either
1795                  * an on-disk directory entry deletion or a bulk data
1796                  * overwrite.  If it is a directory entry deletion we eat
1797                  * both entries.
1798                  *
1799                  * For the bulk-data overwrite case it is possible to have
1800                  * visibility into both, which simply means the syncer
1801                  * hasn't gotten around to doing the delete+insert sequence
1802                  * on the B-Tree.  Use the memory entry and throw away the
1803                  * on-disk entry.
1804                  *
1805                  * If the in-memory record is not either of these we
1806                  * probably caught the syncer while it was syncing it to
1807                  * the media.  Since we hold a shared lock on the cursor,
1808                  * the in-memory record had better be marked deleted at
1809                  * this point.
1810                  */
1811                 if (r == 0) {
1812                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1813                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1814                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1815                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1816                                         goto again;
1817                                 }
1818                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1819                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1820                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1821                                 }
1822                                 /* fall through to memory entry */
1823                         } else {
1824                                 hpanic("duplicate mem/B-Tree entry %p %d %08x",
1825                                         cursor->iprec,
1826                                         cursor->iprec->type,
1827                                         cursor->iprec->flags);
1828                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1829                                 goto again;
1830                         }
1831                 }
1832                 /* fall through to the memory entry */
1833         case HAMMER_CURSOR_ATEDISK:
1834                 /*
1835                  * Only the memory entry is valid.
1836                  */
1837                 cursor->leaf = &cursor->iprec->leaf;
1838                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1839                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1840
1841                 /*
1842                  * If the memory entry is an on-disk deletion we should have
1843                  * also had found a B-Tree record.  If the backend beat us
1844                  * to it it would have interlocked the cursor and we should
1845                  * have seen the in-memory record marked DELETED_FE.
1846                  */
1847                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1848                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1849                         hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x",
1850                                 cursor->iprec,
1851                                 cursor->iprec->flags);
1852                 }
1853                 break;
1854         case HAMMER_CURSOR_ATEMEM:
1855                 /*
1856                  * Only the disk entry is valid
1857                  */
1858                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1859                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1860                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1861                 break;
1862         default:
1863                 /*
1864                  * Neither entry is valid
1865                  *
1866                  * XXX error not set properly
1867                  */
1868                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1869                 cursor->leaf = NULL;
1870                 error = ENOENT;
1871                 break;
1872         }
1873         return(error);
1874 }
1875
1876 /*
1877  * Resolve the cursor->data pointer for the current cursor position in
1878  * a merged iteration.
1879  */
1880 int
1881 hammer_ip_resolve_data(hammer_cursor_t cursor)
1882 {
1883         hammer_record_t record;
1884         int error;
1885
1886         if (hammer_cursor_inmem(cursor)) {
1887                 /*
1888                  * The data associated with an in-memory record is usually
1889                  * kmalloced, but reserve-ahead data records will have an
1890                  * on-disk reference.
1891                  *
1892                  * NOTE: Reserve-ahead data records must be handled in the
1893                  * context of the related high level buffer cache buffer
1894                  * to interlock against async writes.
1895                  */
1896                 record = cursor->iprec;
1897                 cursor->data = record->data;
1898                 error = 0;
1899                 if (cursor->data == NULL) {
1900                         KKASSERT(record->leaf.base.rec_type ==
1901                                  HAMMER_RECTYPE_DATA);
1902                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1903                                                     record->leaf.data_offset,
1904                                                     record->leaf.data_len,
1905                                                     &error,
1906                                                     &cursor->data_buffer);
1907                 }
1908         } else {
1909                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1910                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1911         }
1912         return(error);
1913 }
1914
1915 /*
1916  * Backend truncation / record replacement - delete records in range.
1917  *
1918  * Delete all records within the specified range for inode ip.  In-memory
1919  * records still associated with the frontend are ignored.
1920  *
1921  * If truncating is non-zero in-memory records associated with the back-end
1922  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1923  *
1924  * NOTES:
1925  *
1926  *      * An unaligned range will cause new records to be added to cover
1927  *        the edge cases. (XXX not implemented yet).
1928  *
1929  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1930  *        also do not deal with unaligned ranges.
1931  *
1932  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1933  *
1934  *      * Record keys for regular file data have to be special-cased since
1935  *        they indicate the end of the range (key = base + bytes).
1936  *
1937  *      * This function may be asked to delete ridiculously huge ranges, for
1938  *        example if someone truncates or removes a 1TB regular file.  We
1939  *        must be very careful on restarts and we may have to stop w/
1940  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1941  */
1942 int
1943 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1944                        int64_t ran_beg, int64_t ran_end, int truncating)
1945 {
1946         hammer_transaction_t trans = cursor->trans;
1947         hammer_btree_leaf_elm_t leaf;
1948         int error;
1949         int64_t off;
1950         int64_t tmp64;
1951
1952         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1953 retry:
1954         hammer_normalize_cursor(cursor);
1955         cursor->key_beg.localization = ip->obj_localization |
1956                                        HAMMER_LOCALIZE_MISC;
1957         cursor->key_beg.obj_id = ip->obj_id;
1958         cursor->key_beg.create_tid = 0;
1959         cursor->key_beg.delete_tid = 0;
1960         cursor->key_beg.obj_type = 0;
1961
1962         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1963                 cursor->key_beg.key = ran_beg;
1964                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1965         } else {
1966                 /*
1967                  * The key in the B-Tree is (base+bytes), so the first possible
1968                  * matching key is ran_beg + 1.
1969                  */
1970                 cursor->key_beg.key = ran_beg + 1;
1971                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1972         }
1973
1974         cursor->key_end = cursor->key_beg;
1975         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1976                 cursor->key_end.key = ran_end;
1977         } else {
1978                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1979                 if (tmp64 < ran_end)
1980                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1981                 else
1982                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1983         }
1984
1985         cursor->asof = ip->obj_asof;
1986         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1987         cursor->flags |= HAMMER_CURSOR_ASOF;
1988         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1989         cursor->flags |= HAMMER_CURSOR_BACKEND;
1990         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1991
1992         error = hammer_ip_first(cursor);
1993
1994         /*
1995          * Iterate through matching records and mark them as deleted.
1996          */
1997         while (error == 0) {
1998                 leaf = cursor->leaf;
1999
2000                 KKASSERT(leaf->base.delete_tid == 0);
2001                 KKASSERT(leaf->base.obj_id == ip->obj_id);
2002
2003                 /*
2004                  * There may be overlap cases for regular file data.  Also
2005                  * remember the key for a regular file record is (base + len),
2006                  * NOT (base).
2007                  *
2008                  * Note that due to duplicates (mem & media) allowed by
2009                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
2010                  */
2011                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
2012                         off = leaf->base.key - leaf->data_len;
2013                         /*
2014                          * Check the left edge case.  We currently do not
2015                          * split existing records.
2016                          */
2017                         if (off < ran_beg && leaf->base.key > ran_beg) {
2018                                 hpanic("hammer left edge case %016llx %d",
2019                                         (long long)leaf->base.key,
2020                                         leaf->data_len);
2021                         }
2022
2023                         /*
2024                          * Check the right edge case.  Note that the
2025                          * record can be completely out of bounds, which
2026                          * terminates the search.
2027                          *
2028                          * base->key is exclusive of the right edge while
2029                          * ran_end is inclusive of the right edge.  The
2030                          * (key - data_len) left boundary is inclusive.
2031                          *
2032                          * XXX theory-check this test at some point, are
2033                          * we missing a + 1 somewhere?  Note that ran_end
2034                          * could overflow.
2035                          */
2036                         if (leaf->base.key - 1 > ran_end) {
2037                                 if (leaf->base.key - leaf->data_len > ran_end)
2038                                         break;
2039                                 hpanic("hammer right edge case");
2040                         }
2041                 } else {
2042                         off = leaf->base.key;
2043                 }
2044
2045                 /*
2046                  * Delete the record.  When truncating we do not delete
2047                  * in-memory (data) records because they represent data
2048                  * written after the truncation.
2049                  *
2050                  * This will also physically destroy the B-Tree entry and
2051                  * data if the retention policy dictates.  The function
2052                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2053                  * to retest the new 'current' element.
2054                  */
2055                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2056                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2057                         /*
2058                          * If we have built up too many meta-buffers we risk
2059                          * deadlocking the kernel and must stop.  This can
2060                          * occur when deleting ridiculously huge files.
2061                          * sync_trunc_off is updated so the next cycle does
2062                          * not re-iterate records we have already deleted.
2063                          *
2064                          * This is only done with formal truncations.
2065                          */
2066                         if (truncating > 1 && error == 0 &&
2067                             hammer_flusher_meta_limit(ip->hmp)) {
2068                                 ip->sync_trunc_off = off;
2069                                 error = EWOULDBLOCK;
2070                         }
2071                 }
2072                 if (error)
2073                         break;
2074                 ran_beg = off;  /* for restart */
2075                 error = hammer_ip_next(cursor);
2076         }
2077         if (cursor->node)
2078                 hammer_cache_node(&ip->cache[1], cursor->node);
2079
2080         if (error == EDEADLK) {
2081                 hammer_done_cursor(cursor);
2082                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2083                 if (error == 0)
2084                         goto retry;
2085         }
2086         if (error == ENOENT)
2087                 error = 0;
2088         return(error);
2089 }
2090
2091 /*
2092  * This backend function deletes the specified record on-disk, similar to
2093  * delete_range but for a specific record.  Unlike the exact deletions
2094  * used when deleting a directory entry this function uses an ASOF search
2095  * like delete_range.
2096  *
2097  * This function may be called with ip->obj_asof set for a slave snapshot,
2098  * so don't use it.  We always delete non-historical records only.
2099  */
2100 static int
2101 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2102                       hammer_btree_leaf_elm_t leaf)
2103 {
2104         hammer_transaction_t trans = cursor->trans;
2105         int error;
2106
2107         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2108 retry:
2109         hammer_normalize_cursor(cursor);
2110         cursor->key_beg = leaf->base;
2111         cursor->asof = HAMMER_MAX_TID;
2112         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2113         cursor->flags |= HAMMER_CURSOR_ASOF;
2114         cursor->flags |= HAMMER_CURSOR_BACKEND;
2115         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2116
2117         error = hammer_btree_lookup(cursor);
2118         if (error == 0) {
2119                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2120         }
2121         if (error == EDEADLK) {
2122                 hammer_done_cursor(cursor);
2123                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2124                 if (error == 0)
2125                         goto retry;
2126         }
2127         return(error);
2128 }
2129
2130 /*
2131  * This function deletes remaining auxillary records when an inode is
2132  * being deleted.  This function explicitly does not delete the
2133  * inode record, directory entry, data, or db records.  Those must be
2134  * properly disposed of prior to this call.
2135  */
2136 int
2137 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2138 {
2139         hammer_transaction_t trans = cursor->trans;
2140         hammer_btree_leaf_elm_t leaf __debugvar;
2141         int error;
2142
2143         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2144 retry:
2145         hammer_normalize_cursor(cursor);
2146         cursor->key_beg.localization = ip->obj_localization |
2147                                        HAMMER_LOCALIZE_MISC;
2148         cursor->key_beg.obj_id = ip->obj_id;
2149         cursor->key_beg.create_tid = 0;
2150         cursor->key_beg.delete_tid = 0;
2151         cursor->key_beg.obj_type = 0;
2152         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2153         cursor->key_beg.key = HAMMER_MIN_KEY;
2154
2155         cursor->key_end = cursor->key_beg;
2156         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2157         cursor->key_end.key = HAMMER_MAX_KEY;
2158
2159         cursor->asof = ip->obj_asof;
2160         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2161         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2162         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2163         cursor->flags |= HAMMER_CURSOR_BACKEND;
2164
2165         error = hammer_ip_first(cursor);
2166
2167         /*
2168          * Iterate through matching records and mark them as deleted.
2169          */
2170         while (error == 0) {
2171                 leaf = cursor->leaf;
2172
2173                 KKASSERT(leaf->base.delete_tid == 0);
2174
2175                 /*
2176                  * Mark the record and B-Tree entry as deleted.  This will
2177                  * also physically delete the B-Tree entry, record, and
2178                  * data if the retention policy dictates.  The function
2179                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2180                  * to retest the new 'current' element.
2181                  *
2182                  * Directory entries (and delete-on-disk directory entries)
2183                  * must be synced and cannot be deleted.
2184                  */
2185                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2186                 ++*countp;
2187                 if (error)
2188                         break;
2189                 error = hammer_ip_next(cursor);
2190         }
2191         if (cursor->node)
2192                 hammer_cache_node(&ip->cache[1], cursor->node);
2193         if (error == EDEADLK) {
2194                 hammer_done_cursor(cursor);
2195                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2196                 if (error == 0)
2197                         goto retry;
2198         }
2199         if (error == ENOENT)
2200                 error = 0;
2201         return(error);
2202 }
2203
2204 /*
2205  * Delete the record at the current cursor.  On success the cursor will
2206  * be positioned appropriately for an iteration but may no longer be at
2207  * a leaf node.
2208  *
2209  * This routine is only called from the backend.
2210  *
2211  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2212  * cursor and retry.
2213  */
2214 int
2215 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2216                         hammer_tid_t tid)
2217 {
2218         hammer_record_t iprec;
2219         int error;
2220
2221         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2222         KKASSERT(tid != 0);
2223
2224         /*
2225          * In-memory (unsynchronized) records can simply be freed.  This
2226          * only occurs in range iterations since all other records are
2227          * individually synchronized.  Thus there should be no confusion with
2228          * the interlock.
2229          *
2230          * An in-memory record may be deleted before being committed to disk,
2231          * but could have been accessed in the mean time.  The reservation
2232          * code will deal with the case.
2233          */
2234         if (hammer_cursor_inmem(cursor)) {
2235                 iprec = cursor->iprec;
2236                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2237                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2238                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2239                 KKASSERT(iprec->ip == ip);
2240                 ++ip->rec_generation;
2241                 return(0);
2242         }
2243
2244         /*
2245          * On-disk records are marked as deleted by updating their delete_tid.
2246          * This does not effect their position in the B-Tree (which is based
2247          * on their create_tid).
2248          *
2249          * Frontend B-Tree operations track inodes so we tell
2250          * hammer_delete_at_cursor() not to.
2251          */
2252         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2253
2254         if (error == 0) {
2255                 error = hammer_delete_at_cursor(
2256                                 cursor,
2257                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2258                                 cursor->trans->tid,
2259                                 cursor->trans->time32,
2260                                 0, NULL);
2261         }
2262         return(error);
2263 }
2264
2265 /*
2266  * Used to write a generic record w/optional data to the media b-tree
2267  * when no inode context is available.  Used by the mirroring and
2268  * snapshot code.
2269  *
2270  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2271  * flagged for backend operation and not flagged ASOF (since we are
2272  * doing an insertion).
2273  *
2274  * This function will acquire the appropriate sync lock and will set
2275  * the cursor insertion flag for the operation, do the btree lookup,
2276  * and the insertion, and clear the insertion flag and sync lock before
2277  * returning.  The cursor state will be such that the caller can continue
2278  * scanning (used by the mirroring code).
2279  *
2280  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2281  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2282  *
2283  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2284  *                retry.
2285  *
2286  *       EALREADY can be returned if the record already exists (WARNING,
2287  *                because ASOF cannot be used no check is made for illegal
2288  *                duplicates).
2289  *
2290  * NOTE: Do not use the function for normal inode-related records as this
2291  *       functions goes directly to the media and is not integrated with
2292  *       in-memory records.
2293  */
2294 int
2295 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2296                         void *udata, int mode)
2297 {
2298         hammer_transaction_t trans;
2299         hammer_buffer_t data_buffer;
2300         hammer_off_t ndata_offset;
2301         hammer_tid_t high_tid;
2302         void *ndata;
2303         int error;
2304         int doprop;
2305
2306         trans = cursor->trans;
2307         data_buffer = NULL;
2308         ndata_offset = 0;
2309         doprop = 0;
2310
2311         KKASSERT((cursor->flags &
2312                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2313                   (HAMMER_CURSOR_BACKEND));
2314
2315         hammer_sync_lock_sh(trans);
2316
2317         if (leaf->data_len) {
2318                 ndata = hammer_alloc_data(trans, leaf->data_len,
2319                                           leaf->base.rec_type,
2320                                           &ndata_offset, &data_buffer,
2321                                           0, &error);
2322                 if (ndata == NULL) {
2323                         hammer_sync_unlock(trans);
2324                         return (error);
2325                 }
2326                 leaf->data_offset = ndata_offset;
2327                 hammer_modify_buffer_noundo(trans, data_buffer);
2328
2329                 switch(mode) {
2330                 case HAMMER_CREATE_MODE_UMIRROR:
2331                         error = copyin(udata, ndata, leaf->data_len);
2332                         if (error == 0) {
2333                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2334                                         hdkprintf("CRC DATA @ %016llx/%d MISMATCH ON PIPE\n",
2335                                                 (long long)ndata_offset,
2336                                                 leaf->data_len);
2337                                         error = EINVAL;
2338                                 } else {
2339                                         error = hammer_cursor_localize_data(
2340                                                         ndata, leaf);
2341                                 }
2342                         }
2343                         break;
2344                 case HAMMER_CREATE_MODE_SYS:
2345                         bcopy(udata, ndata, leaf->data_len);
2346                         error = 0;
2347                         hammer_crc_set_leaf(ndata, leaf);
2348                         break;
2349                 default:
2350                         hpanic("bad mode %d", mode);
2351                         break; /* NOT REACHED */
2352                 }
2353                 hammer_modify_buffer_done(data_buffer);
2354         } else {
2355                 leaf->data_offset = 0;
2356                 error = 0;
2357                 ndata = NULL;
2358         }
2359         if (error)
2360                 goto failed;
2361
2362         /*
2363          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2364          */
2365         cursor->flags |= HAMMER_CURSOR_INSERT;
2366         error = hammer_btree_lookup(cursor);
2367         if (error != ENOENT) {
2368                 if (error == 0)
2369                         error = EALREADY;
2370                 goto failed;
2371         }
2372         error = hammer_btree_insert(cursor, leaf, &doprop);
2373
2374         /*
2375          * Cursor is left on current element, we want to skip it now.
2376          * (in case the caller is scanning)
2377          */
2378         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2379         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2380
2381         /*
2382          * If the insertion happens to be creating (and not just replacing)
2383          * an inode we have to track it.
2384          */
2385         if (error == 0 &&
2386             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2387             leaf->base.delete_tid == 0) {
2388                 hammer_modify_volume_field(trans, trans->rootvol,
2389                                            vol0_stat_inodes);
2390                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2391                 hammer_modify_volume_done(trans->rootvol);
2392         }
2393
2394         /*
2395          * vol0_next_tid must track the highest TID stored in the filesystem.
2396          * We do not need to generate undo for this update.
2397          */
2398         high_tid = leaf->base.create_tid;
2399         if (high_tid < leaf->base.delete_tid)
2400                 high_tid = leaf->base.delete_tid;
2401         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2402                 hammer_modify_volume_noundo(trans, trans->rootvol);
2403                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2404                 hammer_modify_volume_done(trans->rootvol);
2405         }
2406
2407         /*
2408          * WARNING!  cursor's leaf pointer may have changed after
2409          *           do_propagation returns.
2410          */
2411         if (error == 0 && doprop)
2412                 hammer_btree_do_propagation(cursor, NULL, leaf);
2413
2414 failed:
2415         /*
2416          * Cleanup
2417          */
2418         if (error && leaf->data_offset) {
2419                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2420
2421         }
2422         hammer_sync_unlock(trans);
2423         if (data_buffer)
2424                 hammer_rel_buffer(data_buffer, 0);
2425         return (error);
2426 }
2427
2428 /*
2429  * Delete the B-Tree element at the current cursor and do any necessary
2430  * mirror propagation.
2431  *
2432  * The cursor must be properly positioned for an iteration on return but
2433  * may be pointing at an internal element.
2434  *
2435  * An element can be un-deleted by passing a delete_tid of 0 with
2436  * HAMMER_DELETE_ADJUST.
2437  *
2438  * This function will store the number of bytes deleted in *stat_bytes
2439  * if stat_bytes is not NULL.
2440  */
2441 int
2442 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2443                         hammer_tid_t delete_tid, uint32_t delete_ts,
2444                         int track, int64_t *stat_bytes)
2445 {
2446         struct hammer_btree_leaf_elm save_leaf;
2447         hammer_transaction_t trans;
2448         hammer_btree_leaf_elm_t leaf;
2449         hammer_node_t node;
2450         hammer_btree_elm_t elm;
2451         hammer_off_t data_offset;
2452         int32_t data_len;
2453         int64_t bytes;
2454         int ndelete;
2455         int error;
2456         int icount;
2457         int doprop;
2458
2459         error = hammer_cursor_upgrade(cursor);
2460         if (error)
2461                 return(error);
2462
2463         trans = cursor->trans;
2464         node = cursor->node;
2465         elm = &node->ondisk->elms[cursor->index];
2466         leaf = &elm->leaf;
2467         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2468
2469         hammer_sync_lock_sh(trans);
2470         bytes = 0;
2471         doprop = 0;
2472         icount = 0;
2473
2474         /*
2475          * Adjust the delete_tid.  Update the mirror_tid propagation field
2476          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2477          */
2478         if (delete_flags & HAMMER_DELETE_ADJUST) {
2479                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2480                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2481                                 icount = -1;
2482                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2483                                 icount = 1;
2484                 }
2485
2486                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2487                 elm->leaf.base.delete_tid = delete_tid;
2488                 elm->leaf.delete_ts = delete_ts;
2489                 hammer_modify_node_done(node);
2490
2491                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2492                         hammer_modify_node_field(trans, node, mirror_tid);
2493                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2494                         hammer_modify_node_done(node);
2495                         doprop = 1;
2496                         if (hammer_debug_general & 0x0002) {
2497                                 hdkprintf("propagate %016llx @%016llx\n",
2498                                         (long long)elm->leaf.base.delete_tid,
2499                                         (long long)node->node_offset);
2500                         }
2501                 }
2502
2503                 /*
2504                  * Adjust for the iteration.  We have deleted the current
2505                  * element and want to clear ATEDISK so the iteration does
2506                  * not skip the element after, which now becomes the current
2507                  * element.  This element must be re-tested if doing an
2508                  * iteration, which is handled by the RETEST flag.
2509                  */
2510                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2511                         cursor->flags |= HAMMER_CURSOR_RETEST;
2512                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2513                 }
2514
2515                 /*
2516                  * An on-disk record cannot have the same delete_tid
2517                  * as its create_tid.  In a chain of record updates
2518                  * this could result in a duplicate record.
2519                  */
2520                 KKASSERT(elm->leaf.base.delete_tid !=
2521                          elm->leaf.base.create_tid);
2522         }
2523
2524         /*
2525          * Destroy the B-Tree element if asked (typically if a nohistory
2526          * file or mount, or when called by the pruning code).
2527          *
2528          * Adjust the ATEDISK flag to properly support iterations.
2529          */
2530         if (delete_flags & HAMMER_DELETE_DESTROY) {
2531                 data_offset = elm->leaf.data_offset;
2532                 data_len = elm->leaf.data_len;
2533                 if (doprop) {
2534                         save_leaf = elm->leaf;
2535                         leaf = &save_leaf;
2536                 }
2537                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2538                     elm->leaf.base.delete_tid == 0) {
2539                         icount = -1;
2540                 }
2541
2542                 error = hammer_btree_delete(cursor, &ndelete);
2543                 if (error == 0) {
2544                         /*
2545                          * The deletion moves the next element (if any) to
2546                          * the current element position.  We must clear
2547                          * ATEDISK so this element is not skipped and we
2548                          * must set RETEST to force any iteration to re-test
2549                          * the element.
2550                          */
2551                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2552                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2553                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2554                         }
2555                         bytes += (ndelete * sizeof(struct hammer_node_ondisk));
2556
2557                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2558                         case HAMMER_ZONE_LARGE_DATA:
2559                         case HAMMER_ZONE_SMALL_DATA:
2560                         case HAMMER_ZONE_META:
2561                                 hammer_blockmap_free(trans,
2562                                                      data_offset, data_len);
2563                                 bytes += data_len;
2564                                 break;
2565                         default:
2566                                 break;
2567                         }
2568                 }
2569         }
2570
2571         /*
2572          * Track inode count and next_tid.  This is used by the mirroring
2573          * and PFS code.  icount can be negative, zero, or positive.
2574          */
2575         if (error == 0 && track) {
2576                 if (icount) {
2577                         hammer_modify_volume_field(trans, trans->rootvol,
2578                                                    vol0_stat_inodes);
2579                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2580                         hammer_modify_volume_done(trans->rootvol);
2581                 }
2582                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2583                         hammer_modify_volume_noundo(trans, trans->rootvol);
2584                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2585                         hammer_modify_volume_done(trans->rootvol);
2586                 }
2587         }
2588
2589         /*
2590          * mirror_tid propagation occurs if the node's mirror_tid had to be
2591          * updated while adjusting the delete_tid.
2592          *
2593          * This occurs when deleting even in nohistory mode, but does not
2594          * occur when pruning an already-deleted node.
2595          *
2596          * cursor->ip is NULL when called from the pruning, mirroring,
2597          * and pfs code.  If non-NULL propagation will be conditionalized
2598          * on whether the PFS is in no-history mode or not.
2599          *
2600          * WARNING: cursor's leaf pointer may have changed after do_propagation
2601          *          returns!
2602          */
2603         if (doprop) {
2604                 if (cursor->ip)
2605                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2606                 else
2607                         hammer_btree_do_propagation(cursor, NULL, leaf);
2608         }
2609         if (stat_bytes)
2610                 *stat_bytes = bytes;
2611         hammer_sync_unlock(trans);
2612         return (error);
2613 }
2614
2615 /*
2616  * Determine whether we can remove a directory.  This routine checks whether
2617  * a directory is empty or not and enforces flush connectivity.
2618  *
2619  * Flush connectivity requires that we block if the target directory is
2620  * currently flushing, otherwise it may not end up in the same flush group.
2621  *
2622  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2623  */
2624 int
2625 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2626 {
2627         struct hammer_cursor cursor;
2628         int error;
2629
2630         /*
2631          * Check directory empty
2632          */
2633         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2634
2635         cursor.key_beg.localization = ip->obj_localization |
2636                                       hammer_dir_localization(ip);
2637         cursor.key_beg.obj_id = ip->obj_id;
2638         cursor.key_beg.create_tid = 0;
2639         cursor.key_beg.delete_tid = 0;
2640         cursor.key_beg.obj_type = 0;
2641         cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START;
2642         cursor.key_beg.key = HAMMER_MIN_KEY;
2643
2644         cursor.key_end = cursor.key_beg;
2645         cursor.key_end.rec_type = HAMMER_RECTYPE_MAX;
2646         cursor.key_end.key = HAMMER_MAX_KEY;
2647
2648         cursor.asof = ip->obj_asof;
2649         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2650
2651         error = hammer_ip_first(&cursor);
2652         if (error == ENOENT)
2653                 error = 0;
2654         else if (error == 0)
2655                 error = ENOTEMPTY;
2656         hammer_done_cursor(&cursor);
2657         return(error);
2658 }
2659
2660 /*
2661  * Localize the data payload.  Directory entries may need their
2662  * localization adjusted.
2663  */
2664 static
2665 int
2666 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2667                             hammer_btree_leaf_elm_t leaf)
2668 {
2669         uint32_t localization;
2670
2671         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2672                 localization = leaf->base.localization &
2673                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2674                 if (data->entry.localization != localization) {
2675                         data->entry.localization = localization;
2676                         hammer_crc_set_leaf(data, leaf);
2677                 }
2678         }
2679         return(0);
2680 }