sys/vfs/hammer/hammer_object.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_object.c,v 1.97 2008/09/23 22:28:56 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_lookup(hammer_cursor_t cursor);
  40 static void hammer_mem_first(hammer_cursor_t cursor);
  41 static int hammer_frontend_trunc_callback(hammer_record_t record,
  42                                 void *data __unused);
  43 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
  46                                 hammer_btree_leaf_elm_t leaf);
  47 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
  48                                 hammer_btree_leaf_elm_t leaf);
  49
  50 struct rec_trunc_info {
  51         u_int16_t       rec_type;
  52         int64_t         trunc_off;
  53 };
  54
  55 struct hammer_bulk_info {
  56         hammer_record_t record;
  57         struct hammer_btree_leaf_elm leaf;
  58 };
  59
  60 /*
  61  * Red-black tree support.  Comparison code for insertion.
  62  */
  63 static int
  64 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  65 {
  66         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  67                 return(-1);
  68         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  69                 return(1);
  70
  71         if (rec1->leaf.base.key < rec2->leaf.base.key)
  72                 return(-1);
  73         if (rec1->leaf.base.key > rec2->leaf.base.key)
  74                 return(1);
  75
  76         /*
  77          * For search & insertion purposes records deleted by the
  78          * frontend or deleted/committed by the backend are silently
  79          * ignored.  Otherwise pipelined insertions will get messed
  80          * up.
  81          *
  82          * rec1 is greater then rec2 if rec1 is marked deleted.
  83          * rec1 is less then rec2 if rec2 is marked deleted.
  84          *
  85          * Multiple deleted records may be present, do not return 0
  86          * if both are marked deleted.
  87          */
  88         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  89                            HAMMER_RECF_COMMITTED)) {
  90                 return(1);
  91         }
  92         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  93                            HAMMER_RECF_COMMITTED)) {
  94                 return(-1);
  95         }
  96
  97         return(0);
  98 }
  99
 100 /*
 101  * Basic record comparison code similar to hammer_btree_cmp().
 102  *
 103  * obj_id is not compared and may not yet be assigned in the record.
 104  */
 105 static int
 106 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 107 {
 108         if (elm->rec_type < rec->leaf.base.rec_type)
 109                 return(-3);
 110         if (elm->rec_type > rec->leaf.base.rec_type)
 111                 return(3);
 112
 113         if (elm->key < rec->leaf.base.key)
 114                 return(-2);
 115         if (elm->key > rec->leaf.base.key)
 116                 return(2);
 117
 118         /*
 119          * Never match against an item deleted by the frontend
 120          * or backend, or committed by the backend.
 121          *
 122          * elm is less then rec if rec is marked deleted.
 123          */
 124         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 125                           HAMMER_RECF_COMMITTED)) {
 126                 return(-1);
 127         }
 128         return(0);
 129 }
 130
 131 /*
 132  * Ranged scan to locate overlapping record(s).  This is used by
 133  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 134  * to use a ranged scan because the keys for data records with the
 135  * same file base offset can be different due to differing data_len's.
 136  *
 137  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 138  */
 139 static int
 140 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 141 {
 142         struct hammer_bulk_info *info = data;
 143         hammer_btree_leaf_elm_t leaf = &info->leaf;
 144
 145         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 146                 return(-3);
 147         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 148                 return(3);
 149
 150         /*
 151          * Overlap compare
 152          */
 153         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 154                 /* rec_beg >= leaf_end */
 155                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 156                         return(2);
 157                 /* rec_end <= leaf_beg */
 158                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 159                         return(-2);
 160         } else {
 161                 if (rec->leaf.base.key < leaf->base.key)
 162                         return(-2);
 163                 if (rec->leaf.base.key > leaf->base.key)
 164                         return(2);
 165         }
 166
 167         /*
 168          * We have to return 0 at this point, even if DELETED_FE is set,
 169          * because returning anything else will cause the scan to ignore
 170          * one of the branches when we really want it to check both.
 171          */
 172         return(0);
 173 }
 174
 175 /*
 176  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 177  * is reversed so the comparison result has to be negated.  key_beg and
 178  * key_end are both range-inclusive.
 179  *
 180  * Localized deletions are not cached in-memory.
 181  */
 182 static
 183 int
 184 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 185 {
 186         hammer_cursor_t cursor = data;
 187         int r;
 188
 189         r = hammer_rec_cmp(&cursor->key_beg, rec);
 190         if (r > 1)
 191                 return(-1);
 192         r = hammer_rec_cmp(&cursor->key_end, rec);
 193         if (r < -1)
 194                 return(1);
 195         return(0);
 196 }
 197
 198 /*
 199  * This compare function is used when simply looking up key_beg.
 200  */
 201 static
 202 int
 203 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 204 {
 205         hammer_cursor_t cursor = data;
 206         int r;
 207
 208         r = hammer_rec_cmp(&cursor->key_beg, rec);
 209         if (r > 1)
 210                 return(-1);
 211         if (r < -1)
 212                 return(1);
 213         return(0);
 214 }
 215
 216 /*
 217  * Locate blocks within the truncation range.  Partial blocks do not count.
 218  */
 219 static
 220 int
 221 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 222 {
 223         struct rec_trunc_info *info = data;
 224
 225         if (rec->leaf.base.rec_type < info->rec_type)
 226                 return(-1);
 227         if (rec->leaf.base.rec_type > info->rec_type)
 228                 return(1);
 229
 230         switch(rec->leaf.base.rec_type) {
 231         case HAMMER_RECTYPE_DB:
 232                 /*
 233                  * DB record key is not beyond the truncation point, retain.
 234                  */
 235                 if (rec->leaf.base.key < info->trunc_off)
 236                         return(-1);
 237                 break;
 238         case HAMMER_RECTYPE_DATA:
 239                 /*
 240                  * DATA record offset start is not beyond the truncation point,
 241                  * retain.
 242                  */
 243                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 244                         return(-1);
 245                 break;
 246         default:
 247                 panic("hammer_rec_trunc_cmp: unexpected record type");
 248         }
 249
 250         /*
 251          * The record start is >= the truncation point, return match,
 252          * the record should be destroyed.
 253          */
 254         return(0);
 255 }
 256
 257 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 258
 259 /*
 260  * Allocate a record for the caller to finish filling in.  The record is
 261  * returned referenced.
 262  */
 263 hammer_record_t
 264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 265 {
 266         hammer_record_t record;
 267         hammer_mount_t hmp;
 268
 269         hmp = ip->hmp;
 270         ++hammer_count_records;
 271         record = kmalloc(sizeof(*record), hmp->m_misc,
 272                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 273         record->flush_state = HAMMER_FST_IDLE;
 274         record->ip = ip;
 275         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 276         record->leaf.data_len = data_len;
 277         hammer_ref(&record->lock);
 278
 279         if (data_len) {
 280                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
 281                 record->flags |= HAMMER_RECF_ALLOCDATA;
 282                 ++hammer_count_record_datas;
 283         }
 284
 285         return (record);
 286 }
 287
 288 void
 289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 290 {
 291         while (record->flush_state == HAMMER_FST_FLUSH) {
 292                 record->flags |= HAMMER_RECF_WANTED;
 293                 tsleep(record, 0, ident, 0);
 294         }
 295 }
 296
 297 /*
 298  * Called from the backend, hammer_inode.c, after a record has been
 299  * flushed to disk.  The record has been exclusively locked by the
 300  * caller and interlocked with BE.
 301  *
 302  * We clean up the state, unlock, and release the record (the record
 303  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 304  */
 305 void
 306 hammer_flush_record_done(hammer_record_t record, int error)
 307 {
 308         hammer_inode_t target_ip;
 309
 310         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 311         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 312
 313         /*
 314          * If an error occured, the backend was unable to sync the
 315          * record to its media.  Leave the record intact.
 316          */
 317         if (error) {
 318                 hammer_critical_error(record->ip->hmp, record->ip, error,
 319                                       "while flushing record");
 320         }
 321
 322         --record->flush_group->refs;
 323         record->flush_group = NULL;
 324
 325         /*
 326          * Adjust the flush state and dependancy based on success or
 327          * failure.
 328          */
 329         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
 330                 if ((target_ip = record->target_ip) != NULL) {
 331                         TAILQ_REMOVE(&target_ip->target_list, record,
 332                                      target_entry);
 333                         record->target_ip = NULL;
 334                         hammer_test_inode(target_ip);
 335                 }
 336                 record->flush_state = HAMMER_FST_IDLE;
 337         } else {
 338                 if (record->target_ip) {
 339                         record->flush_state = HAMMER_FST_SETUP;
 340                         hammer_test_inode(record->ip);
 341                         hammer_test_inode(record->target_ip);
 342                 } else {
 343                         record->flush_state = HAMMER_FST_IDLE;
 344                 }
 345         }
 346         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 347
 348         /*
 349          * Cleanup
 350          */
 351         if (record->flags & HAMMER_RECF_WANTED) {
 352                 record->flags &= ~HAMMER_RECF_WANTED;
 353                 wakeup(record);
 354         }
 355         hammer_rel_mem_record(record);
 356 }
 357
 358 /*
 359  * Release a memory record.  Records marked for deletion are immediately
 360  * removed from the RB-Tree but otherwise left intact until the last ref
 361  * goes away.
 362  */
 363 void
 364 hammer_rel_mem_record(struct hammer_record *record)
 365 {
 366         hammer_mount_t hmp;
 367         hammer_reserve_t resv;
 368         hammer_inode_t ip;
 369         hammer_inode_t target_ip;
 370         int diddrop;
 371
 372         hammer_unref(&record->lock);
 373
 374         if (record->lock.refs == 0) {
 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(ip->lock.refs > 0);
 395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 396
 397                         /*
 398                          * target_ip may have zero refs, we have to ref it
 399                          * to prevent it from being ripped out from under
 400                          * us.
 401                          */
 402                         if ((target_ip = record->target_ip) != NULL) {
 403                                 TAILQ_REMOVE(&target_ip->target_list,
 404                                              record, target_entry);
 405                                 record->target_ip = NULL;
 406                                 hammer_ref(&target_ip->lock);
 407                         }
 408
 409                         /*
 410                          * Remove the record from the B-Tree
 411                          */
 412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 413                                 RB_REMOVE(hammer_rec_rb_tree,
 414                                           &record->ip->rec_tree,
 415                                           record);
 416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 417                                 KKASSERT(ip->rsv_recs > 0);
 418                                 diddrop = 1;
 419                         } else {
 420                                 diddrop = 0;
 421                         }
 422
 423                         /*
 424                          * We must wait for any direct-IO to complete before
 425                          * we can destroy the record because the bio may
 426                          * have a reference to it.
 427                          */
 428                         if (record->flags &
 429                            (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL)) {
 430                                 hammer_io_direct_wait(record);
 431                         }
 432
 433                         /*
 434                          * Account for the completion after the direct IO
 435                          * has completed.
 436                          */
 437                         if (diddrop) {
 438                                 --hmp->rsv_recs;
 439                                 --ip->rsv_recs;
 440                                 hmp->rsv_databytes -= record->leaf.data_len;
 441
 442                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 443                                         record->ip->flags &= ~HAMMER_INODE_XDIRTY;
 444                                         record->ip->sync_flags &= ~HAMMER_INODE_XDIRTY;
 445                                         hammer_test_inode(record->ip);
 446                                 }
 447                                 if (ip->rsv_recs == hammer_limit_inode_recs - 1)
 448                                         wakeup(&ip->rsv_recs);
 449                         }
 450
 451                         /*
 452                          * Do this test after removing record from the B-Tree.
 453                          */
 454                         if (target_ip) {
 455                                 hammer_test_inode(target_ip);
 456                                 hammer_rel_inode(target_ip, 0);
 457                         }
 458
 459                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 460                                 --hammer_count_record_datas;
 461                                 kfree(record->data, hmp->m_misc);
 462                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 463                         }
 464
 465                         /*
 466                          * Release the reservation.
 467                          *
 468                          * If the record was not committed we can theoretically
 469                          * undo the reservation.  However, doing so might
 470                          * create weird edge cases with the ordering of
 471                          * direct writes because the related buffer cache
 472                          * elements are per-vnode.  So we don't try.
 473                          */
 474                         if ((resv = record->resv) != NULL) {
 475                                 /* XXX undo leaf.data_offset,leaf.data_len */
 476                                 hammer_blockmap_reserve_complete(hmp, resv);
 477                                 record->resv = NULL;
 478                         }
 479                         record->data = NULL;
 480                         --hammer_count_records;
 481                         kfree(record, hmp->m_misc);
 482                 }
 483         }
 484 }
 485
 486 /*
 487  * Record visibility depends on whether the record is being accessed by
 488  * the backend or the frontend.  Backend tests ignore the frontend delete
 489  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 490  * must also check for commit races.
 491  *
 492  * Return non-zero if the record is visible, zero if it isn't or if it is
 493  * deleted.  Returns 0 if the record has been comitted (unless the special
 494  * delete-visibility flag is set).  A committed record must be located
 495  * via the media B-Tree.  Returns non-zero if the record is good.
 496  *
 497  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 498  * records to be returned.  This is so pending deletions are detected
 499  * when using an iterator to locate an unused hash key, or when we need
 500  * to locate historical records on-disk to destroy.
 501  */
 502 static __inline
 503 int
 504 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 505 {
 506         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 507                 return(1);
 508         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 509                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 510                                      HAMMER_RECF_COMMITTED)) {
 511                         return(0);
 512                 }
 513         } else {
 514                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 515                                      HAMMER_RECF_DELETED_BE |
 516                                      HAMMER_RECF_COMMITTED)) {
 517                         return(0);
 518                 }
 519         }
 520         return(1);
 521 }
 522
 523 /*
 524  * This callback is used as part of the RB_SCAN function for in-memory
 525  * records.  We terminate it (return -1) as soon as we get a match.
 526  *
 527  * This routine is used by frontend code.
 528  *
 529  * The primary compare code does not account for ASOF lookups.  This
 530  * code handles that case as well as a few others.
 531  */
 532 static
 533 int
 534 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 535 {
 536         hammer_cursor_t cursor = data;
 537
 538         /*
 539          * We terminate on success, so this should be NULL on entry.
 540          */
 541         KKASSERT(cursor->iprec == NULL);
 542
 543         /*
 544          * Skip if the record was marked deleted or committed.
 545          */
 546         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 547                 return(0);
 548
 549         /*
 550          * Skip if not visible due to our as-of TID
 551          */
 552         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 553                 if (cursor->asof < rec->leaf.base.create_tid)
 554                         return(0);
 555                 if (rec->leaf.base.delete_tid &&
 556                     cursor->asof >= rec->leaf.base.delete_tid) {
 557                         return(0);
 558                 }
 559         }
 560
 561         /*
 562          * ref the record.  The record is protected from backend B-Tree
 563          * interactions by virtue of the cursor's IP lock.
 564          */
 565         hammer_ref(&rec->lock);
 566
 567         /*
 568          * The record may have been deleted or committed while we
 569          * were blocked.  XXX remove?
 570          */
 571         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 572                 hammer_rel_mem_record(rec);
 573                 return(0);
 574         }
 575
 576         /*
 577          * Set the matching record and stop the scan.
 578          */
 579         cursor->iprec = rec;
 580         return(-1);
 581 }
 582
 583
 584 /*
 585  * Lookup an in-memory record given the key specified in the cursor.  Works
 586  * just like hammer_btree_lookup() but operates on an inode's in-memory
 587  * record list.
 588  *
 589  * The lookup must fail if the record is marked for deferred deletion.
 590  *
 591  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 592  */
 593 static
 594 int
 595 hammer_mem_lookup(hammer_cursor_t cursor)
 596 {
 597         KKASSERT(cursor->ip);
 598         if (cursor->iprec) {
 599                 hammer_rel_mem_record(cursor->iprec);
 600                 cursor->iprec = NULL;
 601         }
 602         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 603                                    hammer_rec_scan_callback, cursor);
 604
 605         return (cursor->iprec ? 0 : ENOENT);
 606 }
 607
 608 /*
 609  * hammer_mem_first() - locate the first in-memory record matching the
 610  * cursor within the bounds of the key range.
 611  *
 612  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 613  * will set ATEMEM the same as MEMEOF, and does not return any error.
 614  */
 615 static
 616 void
 617 hammer_mem_first(hammer_cursor_t cursor)
 618 {
 619         hammer_inode_t ip;
 620
 621         ip = cursor->ip;
 622         KKASSERT(ip != NULL);
 623
 624         if (cursor->iprec) {
 625                 hammer_rel_mem_record(cursor->iprec);
 626                 cursor->iprec = NULL;
 627         }
 628         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 629                                    hammer_rec_scan_callback, cursor);
 630
 631         if (cursor->iprec)
 632                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
 633         else
 634                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
 635 }
 636
 637 /************************************************************************
 638  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 639  ************************************************************************
 640  *
 641  * These functions manipulate in-memory records.  Such records typically
 642  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 643  */
 644
 645 /*
 646  * Add a directory entry (dip,ncp) which references inode (ip).
 647  *
 648  * Note that the low 32 bits of the namekey are set temporarily to create
 649  * a unique in-memory record, and may be modified a second time when the
 650  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 651  * all 0's when synching to disk, which is not handled here.
 652  *
 653  * NOTE: bytes does not include any terminating \0 on name, and name might
 654  * not be terminated.
 655  */
 656 int
 657 hammer_ip_add_directory(struct hammer_transaction *trans,
 658                      struct hammer_inode *dip, const char *name, int bytes,
 659                      struct hammer_inode *ip)
 660 {
 661         struct hammer_cursor cursor;
 662         hammer_record_t record;
 663         int error;
 664         u_int32_t max_iterations;
 665
 666         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 667
 668         record->type = HAMMER_MEM_RECORD_ADD;
 669         record->leaf.base.localization = dip->obj_localization +
 670                                          hammer_dir_localization(dip);
 671         record->leaf.base.obj_id = dip->obj_id;
 672         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 673                                                          &max_iterations);
 674         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 675         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 676         record->data->entry.obj_id = ip->obj_id;
 677         record->data->entry.localization = ip->obj_localization;
 678         bcopy(name, record->data->entry.name, bytes);
 679
 680         ++ip->ino_data.nlinks;
 681         ip->ino_data.ctime = trans->time;
 682         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 683
 684         /*
 685          * Find an unused namekey.  Both the in-memory record tree and
 686          * the B-Tree are checked.  We do not want historically deleted
 687          * names to create a collision as our iteration space may be limited,
 688          * and since create_tid wouldn't match anyway an ASOF search
 689          * must be used to locate collisions.
 690          *
 691          * delete-visibility is set so pending deletions do not give us
 692          * a false-negative on our ability to use an iterator.
 693          *
 694          * The iterator must not rollover the key.  Directory keys only
 695          * use the positive key space.
 696          */
 697         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 698         cursor.key_beg = record->leaf.base;
 699         cursor.flags |= HAMMER_CURSOR_ASOF;
 700         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 701         cursor.asof = ip->obj_asof;
 702
 703         while (hammer_ip_lookup(&cursor) == 0) {
 704                 ++record->leaf.base.key;
 705                 KKASSERT(record->leaf.base.key > 0);
 706                 cursor.key_beg.key = record->leaf.base.key;
 707                 if (--max_iterations == 0) {
 708                         hammer_rel_mem_record(record);
 709                         error = ENOSPC;
 710                         goto failed;
 711                 }
 712         }
 713
 714         /*
 715          * The target inode and the directory entry are bound together.
 716          */
 717         record->target_ip = ip;
 718         record->flush_state = HAMMER_FST_SETUP;
 719         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 720
 721         /*
 722          * The inode now has a dependancy and must be taken out of the idle
 723          * state.  An inode not in an idle state is given an extra reference.
 724          *
 725          * When transitioning to a SETUP state flag for an automatic reflush
 726          * when the dependancies are disposed of if someone is waiting on
 727          * the inode.
 728          */
 729         if (ip->flush_state == HAMMER_FST_IDLE) {
 730                 hammer_ref(&ip->lock);
 731                 ip->flush_state = HAMMER_FST_SETUP;
 732                 if (ip->flags & HAMMER_INODE_FLUSHW)
 733                         ip->flags |= HAMMER_INODE_REFLUSH;
 734         }
 735         error = hammer_mem_add(record);
 736         if (error == 0) {
 737                 dip->ino_data.mtime = trans->time;
 738                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 739         }
 740 failed:
 741         hammer_done_cursor(&cursor);
 742         return(error);
 743 }
 744
 745 /*
 746  * Delete the directory entry and update the inode link count.  The
 747  * cursor must be seeked to the directory entry record being deleted.
 748  *
 749  * The related inode should be share-locked by the caller.  The caller is
 750  * on the frontend.  It could also be NULL indicating that the directory
 751  * entry being removed has no related inode.
 752  *
 753  * This function can return EDEADLK requiring the caller to terminate
 754  * the cursor, any locks, wait on the returned record, and retry.
 755  */
 756 int
 757 hammer_ip_del_directory(struct hammer_transaction *trans,
 758                      hammer_cursor_t cursor, struct hammer_inode *dip,
 759                      struct hammer_inode *ip)
 760 {
 761         hammer_record_t record;
 762         int error;
 763
 764         if (hammer_cursor_inmem(cursor)) {
 765                 /*
 766                  * In-memory (unsynchronized) records can simply be freed.
 767                  *
 768                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 769                  * by the backend, we must still avoid races against the
 770                  * backend potentially syncing the record to the media.
 771                  *
 772                  * We cannot call hammer_ip_delete_record(), that routine may
 773                  * only be called from the backend.
 774                  */
 775                 record = cursor->iprec;
 776                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 777                                      HAMMER_RECF_DELETED_BE |
 778                                      HAMMER_RECF_COMMITTED)) {
 779                         KKASSERT(cursor->deadlk_rec == NULL);
 780                         hammer_ref(&record->lock);
 781                         cursor->deadlk_rec = record;
 782                         error = EDEADLK;
 783                 } else {
 784                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 785                         record->flags |= HAMMER_RECF_DELETED_FE;
 786                         error = 0;
 787                 }
 788         } else {
 789                 /*
 790                  * If the record is on-disk we have to queue the deletion by
 791                  * the record's key.  This also causes lookups to skip the
 792                  * record (lookups for the purposes of finding an unused
 793                  * directory key do not skip the record).
 794                  */
 795                 KKASSERT(dip->flags &
 796                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 797                 record = hammer_alloc_mem_record(dip, 0);
 798                 record->type = HAMMER_MEM_RECORD_DEL;
 799                 record->leaf.base = cursor->leaf->base;
 800                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
 801
 802                 /*
 803                  * ip may be NULL, indicating the deletion of a directory
 804                  * entry which has no related inode.
 805                  */
 806                 record->target_ip = ip;
 807                 if (ip) {
 808                         record->flush_state = HAMMER_FST_SETUP;
 809                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 810                                           target_entry);
 811                 } else {
 812                         record->flush_state = HAMMER_FST_IDLE;
 813                 }
 814
 815                 /*
 816                  * The inode now has a dependancy and must be taken out of
 817                  * the idle state.  An inode not in an idle state is given
 818                  * an extra reference.
 819                  *
 820                  * When transitioning to a SETUP state flag for an automatic
 821                  * reflush when the dependancies are disposed of if someone
 822                  * is waiting on the inode.
 823                  */
 824                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 825                         hammer_ref(&ip->lock);
 826                         ip->flush_state = HAMMER_FST_SETUP;
 827                         if (ip->flags & HAMMER_INODE_FLUSHW)
 828                                 ip->flags |= HAMMER_INODE_REFLUSH;
 829                 }
 830
 831                 error = hammer_mem_add(record);
 832         }
 833
 834         /*
 835          * One less link.  The file may still be open in the OS even after
 836          * all links have gone away.
 837          *
 838          * We have to terminate the cursor before syncing the inode to
 839          * avoid deadlocking against ourselves.  XXX this may no longer
 840          * be true.
 841          *
 842          * If nlinks drops to zero and the vnode is inactive (or there is
 843          * no vnode), call hammer_inode_unloadable_check() to zonk the
 844          * inode.  If we don't do this here the inode will not be destroyed
 845          * on-media until we unmount.
 846          */
 847         if (error == 0) {
 848                 if (ip) {
 849                         --ip->ino_data.nlinks;  /* do before we might block */
 850                         ip->ino_data.ctime = trans->time;
 851                 }
 852                 dip->ino_data.mtime = trans->time;
 853                 hammer_modify_inode(dip, HAMMER_INODE_MTIME);
 854                 if (ip) {
 855                         hammer_modify_inode(ip, HAMMER_INODE_DDIRTY);
 856                         if (ip->ino_data.nlinks == 0 &&
 857                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 858                                 hammer_done_cursor(cursor);
 859                                 hammer_inode_unloadable_check(ip, 1);
 860                                 hammer_flush_inode(ip, 0);
 861                         }
 862                 }
 863
 864         }
 865         return(error);
 866 }
 867
 868 /*
 869  * Add a record to an inode.
 870  *
 871  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 872  * initialize the following additional fields:
 873  *
 874  * The related inode should be share-locked by the caller.  The caller is
 875  * on the frontend.
 876  *
 877  * record->rec.entry.base.base.key
 878  * record->rec.entry.base.base.rec_type
 879  * record->rec.entry.base.base.data_len
 880  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 881  */
 882 int
 883 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 884 {
 885         hammer_inode_t ip = record->ip;
 886         int error;
 887
 888         KKASSERT(record->leaf.base.localization != 0);
 889         record->leaf.base.obj_id = ip->obj_id;
 890         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 891         error = hammer_mem_add(record);
 892         return(error);
 893 }
 894
 895 /*
 896  * Locate a bulk record in-memory.  Bulk records allow disk space to be
 897  * reserved so the front-end can flush large data writes without having
 898  * to queue the BIO to the flusher.  Only the related record gets queued
 899  * to the flusher.
 900  */
 901
 902 static hammer_record_t
 903 hammer_ip_get_bulk(hammer_inode_t ip, off_t file_offset, int bytes)
 904 {
 905         struct hammer_bulk_info info;
 906
 907         bzero(&info, sizeof(info));
 908         info.leaf.base.obj_id = ip->obj_id;
 909         info.leaf.base.key = file_offset + bytes;
 910         info.leaf.base.create_tid = 0;
 911         info.leaf.base.delete_tid = 0;
 912         info.leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 913         info.leaf.base.obj_type = 0;                            /* unused */
 914         info.leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;        /* unused */
 915         info.leaf.base.localization = ip->obj_localization +    /* unused */
 916                                       HAMMER_LOCALIZE_MISC;
 917         info.leaf.data_len = bytes;
 918
 919         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 920                                    hammer_bulk_scan_callback, &info);
 921
 922         return(info.record);    /* may be NULL */
 923 }
 924
 925 /*
 926  * Take records vetted by overlap_cmp.  The first non-deleted record
 927  * (if any) stops the scan.
 928  */
 929 static int
 930 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 931 {
 932         struct hammer_bulk_info *info = data;
 933
 934         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 935                              HAMMER_RECF_COMMITTED)) {
 936                 return(0);
 937         }
 938         hammer_ref(&record->lock);
 939         info->record = record;
 940         return(-1);                     /* stop scan */
 941 }
 942
 943 /*
 944  * Reserve blockmap space placemarked with an in-memory record.
 945  *
 946  * This routine is called by the frontend in order to be able to directly
 947  * flush a buffer cache buffer.  The frontend has locked the related buffer
 948  * cache buffers and we should be able to manipulate any overlapping
 949  * in-memory records.
 950  *
 951  * The caller is responsible for adding the returned record.
 952  */
 953 hammer_record_t
 954 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 955                    int *errorp)
 956 {
 957         hammer_record_t record;
 958         hammer_record_t conflict;
 959         int zone;
 960
 961         /*
 962          * Deal with conflicting in-memory records.  We cannot have multiple
 963          * in-memory records for the same base offset without seriously
 964          * confusing the backend, including but not limited to the backend
 965          * issuing delete-create-delete or create-delete-create sequences
 966          * and asserting on the delete_tid being the same as the create_tid.
 967          *
 968          * If we encounter a record with the backend interlock set we cannot
 969          * immediately delete it without confusing the backend.
 970          */
 971         while ((conflict = hammer_ip_get_bulk(ip, file_offset, bytes)) !=NULL) {
 972                 if (conflict->flags & HAMMER_RECF_INTERLOCK_BE) {
 973                         conflict->flags |= HAMMER_RECF_WANTED;
 974                         tsleep(conflict, 0, "hmrrc3", 0);
 975                 } else {
 976                         conflict->flags |= HAMMER_RECF_DELETED_FE;
 977                 }
 978                 hammer_rel_mem_record(conflict);
 979         }
 980
 981         /*
 982          * Create a record to cover the direct write.  This is called with
 983          * the related BIO locked so there should be no possible conflict.
 984          *
 985          * The backend is responsible for finalizing the space reserved in
 986          * this record.
 987          *
 988          * XXX bytes not aligned, depend on the reservation code to
 989          * align the reservation.
 990          */
 991         record = hammer_alloc_mem_record(ip, 0);
 992         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 993                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 994         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 995                                                &record->leaf.data_offset,
 996                                                errorp);
 997         if (record->resv == NULL) {
 998                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 999                 hammer_rel_mem_record(record);
1000                 return(NULL);
1001         }
1002         record->type = HAMMER_MEM_RECORD_DATA;
1003         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
1004         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
1005         record->leaf.base.obj_id = ip->obj_id;
1006         record->leaf.base.key = file_offset + bytes;
1007         record->leaf.base.localization = ip->obj_localization +
1008                                          HAMMER_LOCALIZE_MISC;
1009         record->leaf.data_len = bytes;
1010         hammer_crc_set_leaf(data, &record->leaf);
1011         KKASSERT(*errorp == 0);
1012         return(record);
1013 }
1014
1015 /*
1016  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1017  * and records in a flushing state are handled by the backend.  The vnops
1018  * setattr code will handle the block containing the truncation point.
1019  *
1020  * Partial blocks are not deleted.
1021  */
1022 int
1023 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1024 {
1025         struct rec_trunc_info info;
1026
1027         switch(ip->ino_data.obj_type) {
1028         case HAMMER_OBJTYPE_REGFILE:
1029                 info.rec_type = HAMMER_RECTYPE_DATA;
1030                 break;
1031         case HAMMER_OBJTYPE_DBFILE:
1032                 info.rec_type = HAMMER_RECTYPE_DB;
1033                 break;
1034         default:
1035                 return(EINVAL);
1036         }
1037         info.trunc_off = file_size;
1038         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1039                                    hammer_frontend_trunc_callback, &info);
1040         return(0);
1041 }
1042
1043 static int
1044 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1045 {
1046         if (record->flags & HAMMER_RECF_DELETED_FE)
1047                 return(0);
1048         if (record->flush_state == HAMMER_FST_FLUSH)
1049                 return(0);
1050         KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
1051         hammer_ref(&record->lock);
1052         record->flags |= HAMMER_RECF_DELETED_FE;
1053         hammer_rel_mem_record(record);
1054         return(0);
1055 }
1056
1057 /*
1058  * Return 1 if the caller must check for and delete existing records
1059  * before writing out a new data record.
1060  *
1061  * Return 0 if the caller can just insert the record into the B-Tree without
1062  * checking.
1063  */
1064 static int
1065 hammer_record_needs_overwrite_delete(hammer_record_t record)
1066 {
1067         hammer_inode_t ip = record->ip;
1068         int64_t file_offset;
1069         int r;
1070
1071         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1072                 file_offset = record->leaf.base.key;
1073         else
1074                 file_offset = record->leaf.base.key - record->leaf.data_len;
1075         r = (file_offset < ip->save_trunc_off);
1076         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1077                 if (ip->save_trunc_off <= record->leaf.base.key)
1078                         ip->save_trunc_off = record->leaf.base.key + 1;
1079         } else {
1080                 if (ip->save_trunc_off < record->leaf.base.key)
1081                         ip->save_trunc_off = record->leaf.base.key;
1082         }
1083         return(r);
1084 }
1085
1086 /*
1087  * Backend code.  Sync a record to the media.
1088  */
1089 int
1090 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1091 {
1092         hammer_transaction_t trans = cursor->trans;
1093         int64_t file_offset;
1094         int bytes;
1095         void *bdata;
1096         int error;
1097         int doprop;
1098
1099         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1100         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1101         KKASSERT(record->leaf.base.localization != 0);
1102
1103         /*
1104          * Any direct-write related to the record must complete before we
1105          * can sync the record to the on-disk media.
1106          */
1107         if (record->flags & (HAMMER_RECF_DIRECT_IO | HAMMER_RECF_DIRECT_INVAL))
1108                 hammer_io_direct_wait(record);
1109
1110         /*
1111          * If this is a bulk-data record placemarker there may be an existing
1112          * record on-disk, indicating a data overwrite.  If there is the
1113          * on-disk record must be deleted before we can insert our new record.
1114          *
1115          * We've synthesized this record and do not know what the create_tid
1116          * on-disk is, nor how much data it represents.
1117          *
1118          * Keep in mind that (key) for data records is (base_offset + len),
1119          * not (base_offset).  Also, we only want to get rid of on-disk
1120          * records since we are trying to sync our in-memory record, call
1121          * hammer_ip_delete_range() with truncating set to 1 to make sure
1122          * it skips in-memory records.
1123          *
1124          * It is ok for the lookup to return ENOENT.
1125          *
1126          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1127          * to call hammer_ip_delete_range() or not.  This also means we must
1128          * update sync_trunc_off() as we write.
1129          */
1130         if (record->type == HAMMER_MEM_RECORD_DATA &&
1131             hammer_record_needs_overwrite_delete(record)) {
1132                 file_offset = record->leaf.base.key - record->leaf.data_len;
1133                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1134                         ~HAMMER_BUFMASK;
1135                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1136                 error = hammer_ip_delete_range(
1137                                 cursor, record->ip,
1138                                 file_offset, file_offset + bytes - 1,
1139                                 1);
1140                 if (error && error != ENOENT)
1141                         goto done;
1142         }
1143
1144         /*
1145          * If this is a general record there may be an on-disk version
1146          * that must be deleted before we can insert the new record.
1147          */
1148         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1149                 error = hammer_delete_general(cursor, record->ip,
1150                                               &record->leaf);
1151                 if (error && error != ENOENT)
1152                         goto done;
1153         }
1154
1155         /*
1156          * Setup the cursor.
1157          */
1158         hammer_normalize_cursor(cursor);
1159         cursor->key_beg = record->leaf.base;
1160         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1161         cursor->flags |= HAMMER_CURSOR_BACKEND;
1162         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1163
1164         /*
1165          * Records can wind up on-media before the inode itself is on-media.
1166          * Flag the case.
1167          */
1168         record->ip->flags |= HAMMER_INODE_DONDISK;
1169
1170         /*
1171          * If we are deleting a directory entry an exact match must be
1172          * found on-disk.
1173          */
1174         if (record->type == HAMMER_MEM_RECORD_DEL) {
1175                 error = hammer_btree_lookup(cursor);
1176                 if (error == 0) {
1177                         KKASSERT(cursor->iprec == NULL);
1178                         error = hammer_ip_delete_record(cursor, record->ip,
1179                                                         trans->tid);
1180                         if (error == 0) {
1181                                 record->flags |= HAMMER_RECF_DELETED_BE |
1182                                                  HAMMER_RECF_COMMITTED;
1183                                 ++record->ip->rec_generation;
1184                         }
1185                 }
1186                 goto done;
1187         }
1188
1189         /*
1190          * We are inserting.
1191          *
1192          * Issue a lookup to position the cursor and locate the insertion
1193          * point.  The target key should not exist.  If we are creating a
1194          * directory entry we may have to iterate the low 32 bits of the
1195          * key to find an unused key.
1196          */
1197         hammer_sync_lock_sh(trans);
1198         cursor->flags |= HAMMER_CURSOR_INSERT;
1199         error = hammer_btree_lookup(cursor);
1200         if (hammer_debug_inode)
1201                 kprintf("DOINSERT LOOKUP %d\n", error);
1202         if (error == 0) {
1203                 kprintf("hammer_ip_sync_record: duplicate rec "
1204                         "at (%016llx)\n", (long long)record->leaf.base.key);
1205                 if (hammer_debug_critical)
1206                         Debugger("duplicate record1");
1207                 error = EIO;
1208         }
1209 #if 0
1210         if (record->type == HAMMER_MEM_RECORD_DATA)
1211                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1212                         record->leaf.base.key - record->leaf.data_len,
1213                         record->leaf.data_offset, error);
1214 #endif
1215
1216         if (error != ENOENT)
1217                 goto done_unlock;
1218
1219         /*
1220          * Allocate the record and data.  The result buffers will be
1221          * marked as being modified and further calls to
1222          * hammer_modify_buffer() will result in unneeded UNDO records.
1223          *
1224          * Support zero-fill records (data == NULL and data_len != 0)
1225          */
1226         if (record->type == HAMMER_MEM_RECORD_DATA) {
1227                 /*
1228                  * The data portion of a bulk-data record has already been
1229                  * committed to disk, we need only adjust the layer2
1230                  * statistics in the same transaction as our B-Tree insert.
1231                  */
1232                 KKASSERT(record->leaf.data_offset != 0);
1233                 error = hammer_blockmap_finalize(trans,
1234                                                  record->resv,
1235                                                  record->leaf.data_offset,
1236                                                  record->leaf.data_len);
1237         } else if (record->data && record->leaf.data_len) {
1238                 /*
1239                  * Wholely cached record, with data.  Allocate the data.
1240                  */
1241                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1242                                           record->leaf.base.rec_type,
1243                                           &record->leaf.data_offset,
1244                                           &cursor->data_buffer,
1245                                           0, &error);
1246                 if (bdata == NULL)
1247                         goto done_unlock;
1248                 hammer_crc_set_leaf(record->data, &record->leaf);
1249                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1250                 bcopy(record->data, bdata, record->leaf.data_len);
1251                 hammer_modify_buffer_done(cursor->data_buffer);
1252         } else {
1253                 /*
1254                  * Wholely cached record, without data.
1255                  */
1256                 record->leaf.data_offset = 0;
1257                 record->leaf.data_crc = 0;
1258         }
1259
1260         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1261         if (hammer_debug_inode && error) {
1262                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1263                         error,
1264                         (long long)cursor->node->node_offset,
1265                         cursor->index,
1266                         (long long)record->leaf.base.key);
1267         }
1268
1269         /*
1270          * Our record is on-disk and we normally mark the in-memory version
1271          * as having been committed (and not BE-deleted).
1272          *
1273          * If the record represented a directory deletion but we had to
1274          * sync a valid directory entry to disk due to dependancies,
1275          * we must convert the record to a covering delete so the
1276          * frontend does not have visibility on the synced entry.
1277          *
1278          * WARNING: cursor's leaf pointer may have changed after do_propagation
1279          *          returns!
1280          */
1281         if (error == 0) {
1282                 if (doprop) {
1283                         hammer_btree_do_propagation(cursor,
1284                                                     record->ip->pfsm,
1285                                                     &record->leaf);
1286                 }
1287                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1288                         /*
1289                          * Must convert deleted directory entry add
1290                          * to a directory entry delete.
1291                          */
1292                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1293                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1294                         record->type = HAMMER_MEM_RECORD_DEL;
1295                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1296                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1297                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1298                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1299                                                  HAMMER_RECF_DELETED_BE)) == 0);
1300                         /* converted record is not yet committed */
1301                         /* hammer_flush_record_done takes care of the rest */
1302                 } else {
1303                         /*
1304                          * Everything went fine and we are now done with
1305                          * this record.
1306                          */
1307                         record->flags |= HAMMER_RECF_COMMITTED;
1308                         ++record->ip->rec_generation;
1309                 }
1310         } else {
1311                 if (record->leaf.data_offset) {
1312                         hammer_blockmap_free(trans, record->leaf.data_offset,
1313                                              record->leaf.data_len);
1314                 }
1315         }
1316 done_unlock:
1317         hammer_sync_unlock(trans);
1318 done:
1319         return(error);
1320 }
1321
1322 /*
1323  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1324  * entry's key is used to deal with hash collisions in the upper 32 bits.
1325  * A unique 64 bit key is generated in-memory and may be regenerated a
1326  * second time when the directory record is flushed to the on-disk B-Tree.
1327  *
1328  * A referenced record is passed to this function.  This function
1329  * eats the reference.  If an error occurs the record will be deleted.
1330  *
1331  * A copy of the temporary record->data pointer provided by the caller
1332  * will be made.
1333  */
1334 int
1335 hammer_mem_add(hammer_record_t record)
1336 {
1337         hammer_mount_t hmp = record->ip->hmp;
1338
1339         /*
1340          * Make a private copy of record->data
1341          */
1342         if (record->data)
1343                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1344
1345         /*
1346          * Insert into the RB tree.  A unique key should have already
1347          * been selected if this is a directory entry.
1348          */
1349         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1350                 record->flags |= HAMMER_RECF_DELETED_FE;
1351                 hammer_rel_mem_record(record);
1352                 return (EEXIST);
1353         }
1354         ++hmp->count_newrecords;
1355         ++hmp->rsv_recs;
1356         ++record->ip->rsv_recs;
1357         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1358         record->flags |= HAMMER_RECF_ONRBTREE;
1359         hammer_modify_inode(record->ip, HAMMER_INODE_XDIRTY);
1360         hammer_rel_mem_record(record);
1361         return(0);
1362 }
1363
1364 /************************************************************************
1365  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1366  ************************************************************************
1367  *
1368  * These functions augment the B-Tree scanning functions in hammer_btree.c
1369  * by merging in-memory records with on-disk records.
1370  */
1371
1372 /*
1373  * Locate a particular record either in-memory or on-disk.
1374  *
1375  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1376  * NOT be called to iterate results.
1377  */
1378 int
1379 hammer_ip_lookup(hammer_cursor_t cursor)
1380 {
1381         int error;
1382
1383         /*
1384          * If the element is in-memory return it without searching the
1385          * on-disk B-Tree
1386          */
1387         KKASSERT(cursor->ip);
1388         error = hammer_mem_lookup(cursor);
1389         if (error == 0) {
1390                 cursor->leaf = &cursor->iprec->leaf;
1391                 return(error);
1392         }
1393         if (error != ENOENT)
1394                 return(error);
1395
1396         /*
1397          * If the inode has on-disk components search the on-disk B-Tree.
1398          */
1399         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1400                 return(error);
1401         error = hammer_btree_lookup(cursor);
1402         if (error == 0)
1403                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1404         return(error);
1405 }
1406
1407 /*
1408  * Helper for hammer_ip_first()/hammer_ip_next()
1409  *
1410  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1411  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1412  * state if hammer_ip_next() needs to re-seek.
1413  */
1414 static __inline
1415 int
1416 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1417 {
1418         hammer_inode_t ip = cursor->ip;
1419         int error;
1420
1421         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1422                 error = hammer_btree_lookup(cursor);
1423                 if (error == ENOENT || error == EDEADLK) {
1424                         if (hammer_debug_general & 0x2000) {
1425                                 kprintf("error %d node %p %016llx index %d\n",
1426                                         error, cursor->node,
1427                                         (long long)cursor->node->node_offset,
1428                                         cursor->index);
1429                         }
1430                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1431                         error = hammer_btree_iterate(cursor);
1432                 }
1433                 if (error == 0) {
1434                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1435                                            HAMMER_CURSOR_ATEDISK);
1436                 } else {
1437                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1438                                          HAMMER_CURSOR_ATEDISK;
1439                         if (error == ENOENT)
1440                                 error = 0;
1441                 }
1442         } else {
1443                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1444                 error = 0;
1445         }
1446         return(error);
1447 }
1448
1449 /*
1450  * Helper for hammer_ip_next()
1451  *
1452  * The caller has determined that the media cursor is further along than the
1453  * memory cursor and must be reseeked after a generation number change.
1454  */
1455 static
1456 int
1457 _hammer_ip_reseek(hammer_cursor_t cursor)
1458 {
1459         struct hammer_base_elm save;
1460         hammer_btree_elm_t elm;
1461         int error;
1462         int r;
1463         int again = 0;
1464
1465         /*
1466          * Do the re-seek.
1467          */
1468         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1469                 (long long)cursor->ip->obj_id);
1470         save = cursor->key_beg;
1471         cursor->key_beg = cursor->iprec->leaf.base;
1472         error = _hammer_ip_seek_btree(cursor);
1473         KKASSERT(error == 0);
1474         cursor->key_beg = save;
1475
1476         /*
1477          * If the memory record was previous returned to
1478          * the caller and the media record matches
1479          * (-1/+1: only create_tid differs), then iterate
1480          * the media record to avoid a double result.
1481          */
1482         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1483             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1484                 elm = &cursor->node->ondisk->elms[cursor->index];
1485                 r = hammer_btree_cmp(&elm->base,
1486                                      &cursor->iprec->leaf.base);
1487                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1488                         if (r >= -1 && r <= 1) {
1489                                 kprintf("HAMMER: Debug: iterated after "
1490                                         "re-seek (asof r=%d)\n", r);
1491                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1492                                 again = 1;
1493                         }
1494                 } else {
1495                         if (r == 0) {
1496                                 kprintf("HAMMER: Debug: iterated after "
1497                                         "re-seek\n");
1498                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1499                                 again = 1;
1500                         }
1501                 }
1502         }
1503         return(again);
1504 }
1505
1506 /*
1507  * Locate the first record within the cursor's key_beg/key_end range,
1508  * restricted to a particular inode.  0 is returned on success, ENOENT
1509  * if no records matched the requested range, or some other error.
1510  *
1511  * When 0 is returned hammer_ip_next() may be used to iterate additional
1512  * records within the requested range.
1513  *
1514  * This function can return EDEADLK, requiring the caller to terminate
1515  * the cursor and try again.
1516  */
1517
1518 int
1519 hammer_ip_first(hammer_cursor_t cursor)
1520 {
1521         hammer_inode_t ip __debugvar = cursor->ip;
1522         int error;
1523
1524         KKASSERT(ip != NULL);
1525
1526         /*
1527          * Clean up fields and setup for merged scan
1528          */
1529         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1530
1531         /*
1532          * Search the in-memory record list (Red-Black tree).  Unlike the
1533          * B-Tree search, mem_first checks for records in the range.
1534          *
1535          * This function will setup both ATEMEM and MEMEOF properly for
1536          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1537          */
1538         hammer_mem_first(cursor);
1539
1540         /*
1541          * Detect generation changes during blockages, including
1542          * blockages which occur on the initial btree search.
1543          */
1544         cursor->rec_generation = cursor->ip->rec_generation;
1545
1546         /*
1547          * Initial search and result
1548          */
1549         error = _hammer_ip_seek_btree(cursor);
1550         if (error == 0)
1551                 error = hammer_ip_next(cursor);
1552
1553         return (error);
1554 }
1555
1556 /*
1557  * Retrieve the next record in a merged iteration within the bounds of the
1558  * cursor.  This call may be made multiple times after the cursor has been
1559  * initially searched with hammer_ip_first().
1560  *
1561  * There are numerous special cases in this code to deal with races between
1562  * in-memory records and on-media records.
1563  *
1564  * 0 is returned on success, ENOENT if no further records match the
1565  * requested range, or some other error code is returned.
1566  */
1567 int
1568 hammer_ip_next(hammer_cursor_t cursor)
1569 {
1570         hammer_btree_elm_t elm;
1571         hammer_record_t rec;
1572         hammer_record_t tmprec;
1573         int error;
1574         int r;
1575
1576 again:
1577         /*
1578          * Get the next on-disk record
1579          *
1580          * NOTE: If we deleted the last on-disk record we had scanned
1581          *       ATEDISK will be clear and RETEST will be set, forcing
1582          *       a call to iterate.  The fact that ATEDISK is clear causes
1583          *       iterate to re-test the 'current' element.  If ATEDISK is
1584          *       set, iterate will skip the 'current' element.
1585          */
1586         error = 0;
1587         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1588                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1589                                      HAMMER_CURSOR_RETEST)) {
1590                         error = hammer_btree_iterate(cursor);
1591                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1592                         if (error == 0) {
1593                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1594                                 hammer_cache_node(&cursor->ip->cache[1],
1595                                                   cursor->node);
1596                         } else if (error == ENOENT) {
1597                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1598                                                  HAMMER_CURSOR_ATEDISK;
1599                                 error = 0;
1600                         }
1601                 }
1602         }
1603
1604         /*
1605          * If the generation changed the backend has deleted or committed
1606          * one or more memory records since our last check.
1607          *
1608          * When this case occurs if the disk cursor is > current memory record
1609          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1610          * Since the cursor is ahead it must have not yet been eaten (if
1611          * not at eof anyway). (XXX data offset case?)
1612          *
1613          * NOTE: we are not doing a full check here.  That will be handled
1614          * later on.
1615          *
1616          * If we have exhausted all memory records we do not have to do any
1617          * further seeks.
1618          */
1619         while (cursor->rec_generation != cursor->ip->rec_generation &&
1620                error == 0
1621         ) {
1622                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1623                 cursor->rec_generation = cursor->ip->rec_generation;
1624                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1625                         break;
1626                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1627                         r = 1;
1628                 } else {
1629                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1630                         elm = &cursor->node->ondisk->elms[cursor->index];
1631                         r = hammer_btree_cmp(&elm->base,
1632                                              &cursor->iprec->leaf.base);
1633                 }
1634
1635                 /*
1636                  * Do we re-seek the media cursor?
1637                  */
1638                 if (r > 0) {
1639                         if (_hammer_ip_reseek(cursor))
1640                                 goto again;
1641                 }
1642         }
1643
1644         /*
1645          * We can now safely get the next in-memory record.  We cannot
1646          * block here.
1647          *
1648          * hammer_rec_scan_cmp:  Is the record still in our general range,
1649          *                       (non-inclusive of snapshot exclusions)?
1650          * hammer_rec_scan_callback: Is the record in our snapshot?
1651          */
1652         tmprec = NULL;
1653         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1654                 /*
1655                  * If the current memory record was eaten then get the next
1656                  * one.  Stale records are skipped.
1657                  */
1658                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1659                         tmprec = cursor->iprec;
1660                         cursor->iprec = NULL;
1661                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1662                         while (rec) {
1663                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1664                                         break;
1665                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1666                                         break;
1667                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1668                         }
1669                         if (cursor->iprec) {
1670                                 KKASSERT(cursor->iprec == rec);
1671                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1672                         } else {
1673                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1674                         }
1675                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1676                 }
1677         }
1678
1679         /*
1680          * MEMORY RECORD VALIDITY TEST
1681          *
1682          * (We still can't block, which is why tmprec is being held so
1683          * long).
1684          *
1685          * If the memory record is no longer valid we skip it.  It may
1686          * have been deleted by the frontend.  If it was deleted or
1687          * committed by the backend the generation change re-seeked the
1688          * disk cursor and the record will be present there.
1689          */
1690         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1691                 KKASSERT(cursor->iprec);
1692                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1693                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1694                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1695                         if (tmprec)
1696                                 hammer_rel_mem_record(tmprec);
1697                         goto again;
1698                 }
1699         }
1700         if (tmprec)
1701                 hammer_rel_mem_record(tmprec);
1702
1703         /*
1704          * Extract either the disk or memory record depending on their
1705          * relative position.
1706          */
1707         error = 0;
1708         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1709         case 0:
1710                 /*
1711                  * Both entries valid.   Compare the entries and nominally
1712                  * return the first one in the sort order.  Numerous cases
1713                  * require special attention, however.
1714                  */
1715                 elm = &cursor->node->ondisk->elms[cursor->index];
1716                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1717
1718                 /*
1719                  * If the two entries differ only by their key (-2/2) or
1720                  * create_tid (-1/1), and are DATA records, we may have a
1721                  * nominal match.  We have to calculate the base file
1722                  * offset of the data.
1723                  */
1724                 if (r <= 2 && r >= -2 && r != 0 &&
1725                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1726                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1727                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1728                         int64_t base2 = cursor->iprec->leaf.base.key -
1729                                         cursor->iprec->leaf.data_len;
1730                         if (base1 == base2)
1731                                 r = 0;
1732                 }
1733
1734                 if (r < 0) {
1735                         error = hammer_btree_extract(cursor,
1736                                                      HAMMER_CURSOR_GET_LEAF);
1737                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1738                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1739                         break;
1740                 }
1741
1742                 /*
1743                  * If the entries match exactly the memory entry is either
1744                  * an on-disk directory entry deletion or a bulk data
1745                  * overwrite.  If it is a directory entry deletion we eat
1746                  * both entries.
1747                  *
1748                  * For the bulk-data overwrite case it is possible to have
1749                  * visibility into both, which simply means the syncer
1750                  * hasn't gotten around to doing the delete+insert sequence
1751                  * on the B-Tree.  Use the memory entry and throw away the
1752                  * on-disk entry.
1753                  *
1754                  * If the in-memory record is not either of these we
1755                  * probably caught the syncer while it was syncing it to
1756                  * the media.  Since we hold a shared lock on the cursor,
1757                  * the in-memory record had better be marked deleted at
1758                  * this point.
1759                  */
1760                 if (r == 0) {
1761                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1762                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1763                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1764                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1765                                         goto again;
1766                                 }
1767                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1768                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1769                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1770                                 }
1771                                 /* fall through to memory entry */
1772                         } else {
1773                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1774                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1775                                 goto again;
1776                         }
1777                 }
1778                 /* fall through to the memory entry */
1779         case HAMMER_CURSOR_ATEDISK:
1780                 /*
1781                  * Only the memory entry is valid.
1782                  */
1783                 cursor->leaf = &cursor->iprec->leaf;
1784                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1785                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1786
1787                 /*
1788                  * If the memory entry is an on-disk deletion we should have
1789                  * also had found a B-Tree record.  If the backend beat us
1790                  * to it it would have interlocked the cursor and we should
1791                  * have seen the in-memory record marked DELETED_FE.
1792                  */
1793                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1794                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1795                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1796                 }
1797                 break;
1798         case HAMMER_CURSOR_ATEMEM:
1799                 /*
1800                  * Only the disk entry is valid
1801                  */
1802                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1803                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1804                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1805                 break;
1806         default:
1807                 /*
1808                  * Neither entry is valid
1809                  *
1810                  * XXX error not set properly
1811                  */
1812                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1813                 cursor->leaf = NULL;
1814                 error = ENOENT;
1815                 break;
1816         }
1817         return(error);
1818 }
1819
1820 /*
1821  * Resolve the cursor->data pointer for the current cursor position in
1822  * a merged iteration.
1823  */
1824 int
1825 hammer_ip_resolve_data(hammer_cursor_t cursor)
1826 {
1827         hammer_record_t record;
1828         int error;
1829
1830         if (hammer_cursor_inmem(cursor)) {
1831                 /*
1832                  * The data associated with an in-memory record is usually
1833                  * kmalloced, but reserve-ahead data records will have an
1834                  * on-disk reference.
1835                  *
1836                  * NOTE: Reserve-ahead data records must be handled in the
1837                  * context of the related high level buffer cache buffer
1838                  * to interlock against async writes.
1839                  */
1840                 record = cursor->iprec;
1841                 cursor->data = record->data;
1842                 error = 0;
1843                 if (cursor->data == NULL) {
1844                         KKASSERT(record->leaf.base.rec_type ==
1845                                  HAMMER_RECTYPE_DATA);
1846                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1847                                                     record->leaf.data_offset,
1848                                                     record->leaf.data_len,
1849                                                     &error,
1850                                                     &cursor->data_buffer);
1851                 }
1852         } else {
1853                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1854                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1855         }
1856         return(error);
1857 }
1858
1859 /*
1860  * Backend truncation / record replacement - delete records in range.
1861  *
1862  * Delete all records within the specified range for inode ip.  In-memory
1863  * records still associated with the frontend are ignored.
1864  *
1865  * If truncating is non-zero in-memory records associated with the back-end
1866  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1867  *
1868  * NOTES:
1869  *
1870  *      * An unaligned range will cause new records to be added to cover
1871  *        the edge cases. (XXX not implemented yet).
1872  *
1873  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1874  *        also do not deal with unaligned ranges.
1875  *
1876  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1877  *
1878  *      * Record keys for regular file data have to be special-cased since
1879  *        they indicate the end of the range (key = base + bytes).
1880  *
1881  *      * This function may be asked to delete ridiculously huge ranges, for
1882  *        example if someone truncates or removes a 1TB regular file.  We
1883  *        must be very careful on restarts and we may have to stop w/
1884  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1885  */
1886 int
1887 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1888                        int64_t ran_beg, int64_t ran_end, int truncating)
1889 {
1890         hammer_transaction_t trans = cursor->trans;
1891         hammer_btree_leaf_elm_t leaf;
1892         int error;
1893         int64_t off;
1894         int64_t tmp64;
1895
1896 #if 0
1897         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1898 #endif
1899
1900         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1901 retry:
1902         hammer_normalize_cursor(cursor);
1903         cursor->key_beg.localization = ip->obj_localization +
1904                                        HAMMER_LOCALIZE_MISC;
1905         cursor->key_beg.obj_id = ip->obj_id;
1906         cursor->key_beg.create_tid = 0;
1907         cursor->key_beg.delete_tid = 0;
1908         cursor->key_beg.obj_type = 0;
1909
1910         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1911                 cursor->key_beg.key = ran_beg;
1912                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1913         } else {
1914                 /*
1915                  * The key in the B-Tree is (base+bytes), so the first possible
1916                  * matching key is ran_beg + 1.
1917                  */
1918                 cursor->key_beg.key = ran_beg + 1;
1919                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1920         }
1921
1922         cursor->key_end = cursor->key_beg;
1923         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1924                 cursor->key_end.key = ran_end;
1925         } else {
1926                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1927                 if (tmp64 < ran_end)
1928                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1929                 else
1930                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1931         }
1932
1933         cursor->asof = ip->obj_asof;
1934         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1935         cursor->flags |= HAMMER_CURSOR_ASOF;
1936         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1937         cursor->flags |= HAMMER_CURSOR_BACKEND;
1938         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1939
1940         error = hammer_ip_first(cursor);
1941
1942         /*
1943          * Iterate through matching records and mark them as deleted.
1944          */
1945         while (error == 0) {
1946                 leaf = cursor->leaf;
1947
1948                 KKASSERT(leaf->base.delete_tid == 0);
1949                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1950
1951                 /*
1952                  * There may be overlap cases for regular file data.  Also
1953                  * remember the key for a regular file record is (base + len),
1954                  * NOT (base).
1955                  *
1956                  * Note that do to duplicates (mem & media) allowed by
1957                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1958                  */
1959                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1960                         off = leaf->base.key - leaf->data_len;
1961                         /*
1962                          * Check the left edge case.  We currently do not
1963                          * split existing records.
1964                          */
1965                         if (off < ran_beg && leaf->base.key > ran_beg) {
1966                                 panic("hammer left edge case %016llx %d\n",
1967                                         (long long)leaf->base.key,
1968                                         leaf->data_len);
1969                         }
1970
1971                         /*
1972                          * Check the right edge case.  Note that the
1973                          * record can be completely out of bounds, which
1974                          * terminates the search.
1975                          *
1976                          * base->key is exclusive of the right edge while
1977                          * ran_end is inclusive of the right edge.  The
1978                          * (key - data_len) left boundary is inclusive.
1979                          *
1980                          * XXX theory-check this test at some point, are
1981                          * we missing a + 1 somewhere?  Note that ran_end
1982                          * could overflow.
1983                          */
1984                         if (leaf->base.key - 1 > ran_end) {
1985                                 if (leaf->base.key - leaf->data_len > ran_end)
1986                                         break;
1987                                 panic("hammer right edge case\n");
1988                         }
1989                 } else {
1990                         off = leaf->base.key;
1991                 }
1992
1993                 /*
1994                  * Delete the record.  When truncating we do not delete
1995                  * in-memory (data) records because they represent data
1996                  * written after the truncation.
1997                  *
1998                  * This will also physically destroy the B-Tree entry and
1999                  * data if the retention policy dictates.  The function
2000                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2001                  * to retest the new 'current' element.
2002                  */
2003                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2004                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2005                         /*
2006                          * If we have built up too many meta-buffers we risk
2007                          * deadlocking the kernel and must stop.  This can
2008                          * occur when deleting ridiculously huge files.
2009                          * sync_trunc_off is updated so the next cycle does
2010                          * not re-iterate records we have already deleted.
2011                          *
2012                          * This is only done with formal truncations.
2013                          */
2014                         if (truncating > 1 && error == 0 &&
2015                             hammer_flusher_meta_limit(ip->hmp)) {
2016                                 ip->sync_trunc_off = off;
2017                                 error = EWOULDBLOCK;
2018                         }
2019                 }
2020                 if (error)
2021                         break;
2022                 ran_beg = off;  /* for restart */
2023                 error = hammer_ip_next(cursor);
2024         }
2025         if (cursor->node)
2026                 hammer_cache_node(&ip->cache[1], cursor->node);
2027
2028         if (error == EDEADLK) {
2029                 hammer_done_cursor(cursor);
2030                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2031                 if (error == 0)
2032                         goto retry;
2033         }
2034         if (error == ENOENT)
2035                 error = 0;
2036         return(error);
2037 }
2038
2039 /*
2040  * This backend function deletes the specified record on-disk, similar to
2041  * delete_range but for a specific record.  Unlike the exact deletions
2042  * used when deleting a directory entry this function uses an ASOF search
2043  * like delete_range.
2044  *
2045  * This function may be called with ip->obj_asof set for a slave snapshot,
2046  * so don't use it.  We always delete non-historical records only.
2047  */
2048 static int
2049 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2050                       hammer_btree_leaf_elm_t leaf)
2051 {
2052         hammer_transaction_t trans = cursor->trans;
2053         int error;
2054
2055         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2056 retry:
2057         hammer_normalize_cursor(cursor);
2058         cursor->key_beg = leaf->base;
2059         cursor->asof = HAMMER_MAX_TID;
2060         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2061         cursor->flags |= HAMMER_CURSOR_ASOF;
2062         cursor->flags |= HAMMER_CURSOR_BACKEND;
2063         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2064
2065         error = hammer_btree_lookup(cursor);
2066         if (error == 0) {
2067                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2068         }
2069         if (error == EDEADLK) {
2070                 hammer_done_cursor(cursor);
2071                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2072                 if (error == 0)
2073                         goto retry;
2074         }
2075         return(error);
2076 }
2077
2078 /*
2079  * This function deletes remaining auxillary records when an inode is
2080  * being deleted.  This function explicitly does not delete the
2081  * inode record, directory entry, data, or db records.  Those must be
2082  * properly disposed of prior to this call.
2083  */
2084 int
2085 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2086 {
2087         hammer_transaction_t trans = cursor->trans;
2088         hammer_btree_leaf_elm_t leaf;
2089         int error;
2090
2091         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2092 retry:
2093         hammer_normalize_cursor(cursor);
2094         cursor->key_beg.localization = ip->obj_localization +
2095                                        HAMMER_LOCALIZE_MISC;
2096         cursor->key_beg.obj_id = ip->obj_id;
2097         cursor->key_beg.create_tid = 0;
2098         cursor->key_beg.delete_tid = 0;
2099         cursor->key_beg.obj_type = 0;
2100         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2101         cursor->key_beg.key = HAMMER_MIN_KEY;
2102
2103         cursor->key_end = cursor->key_beg;
2104         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2105         cursor->key_end.key = HAMMER_MAX_KEY;
2106
2107         cursor->asof = ip->obj_asof;
2108         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2109         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2110         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2111         cursor->flags |= HAMMER_CURSOR_BACKEND;
2112
2113         error = hammer_ip_first(cursor);
2114
2115         /*
2116          * Iterate through matching records and mark them as deleted.
2117          */
2118         while (error == 0) {
2119                 leaf = cursor->leaf;
2120
2121                 KKASSERT(leaf->base.delete_tid == 0);
2122
2123                 /*
2124                  * Mark the record and B-Tree entry as deleted.  This will
2125                  * also physically delete the B-Tree entry, record, and
2126                  * data if the retention policy dictates.  The function
2127                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2128                  * to retest the new 'current' element.
2129                  *
2130                  * Directory entries (and delete-on-disk directory entries)
2131                  * must be synced and cannot be deleted.
2132                  */
2133                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2134                 ++*countp;
2135                 if (error)
2136                         break;
2137                 error = hammer_ip_next(cursor);
2138         }
2139         if (cursor->node)
2140                 hammer_cache_node(&ip->cache[1], cursor->node);
2141         if (error == EDEADLK) {
2142                 hammer_done_cursor(cursor);
2143                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2144                 if (error == 0)
2145                         goto retry;
2146         }
2147         if (error == ENOENT)
2148                 error = 0;
2149         return(error);
2150 }
2151
2152 /*
2153  * Delete the record at the current cursor.  On success the cursor will
2154  * be positioned appropriately for an iteration but may no longer be at
2155  * a leaf node.
2156  *
2157  * This routine is only called from the backend.
2158  *
2159  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2160  * cursor and retry.
2161  */
2162 int
2163 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2164                         hammer_tid_t tid)
2165 {
2166         hammer_record_t iprec;
2167         hammer_mount_t hmp;
2168         int error;
2169
2170         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2171         KKASSERT(tid != 0);
2172         hmp = cursor->node->hmp;
2173
2174         /*
2175          * In-memory (unsynchronized) records can simply be freed.  This
2176          * only occurs in range iterations since all other records are
2177          * individually synchronized.  Thus there should be no confusion with
2178          * the interlock.
2179          *
2180          * An in-memory record may be deleted before being committed to disk,
2181          * but could have been accessed in the mean time.  The reservation
2182          * code will deal with the case.
2183          */
2184         if (hammer_cursor_inmem(cursor)) {
2185                 iprec = cursor->iprec;
2186                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2187                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2188                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2189                 KKASSERT(iprec->ip == ip);
2190                 ++ip->rec_generation;
2191                 return(0);
2192         }
2193
2194         /*
2195          * On-disk records are marked as deleted by updating their delete_tid.
2196          * This does not effect their position in the B-Tree (which is based
2197          * on their create_tid).
2198          *
2199          * Frontend B-Tree operations track inodes so we tell
2200          * hammer_delete_at_cursor() not to.
2201          */
2202         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2203
2204         if (error == 0) {
2205                 error = hammer_delete_at_cursor(
2206                                 cursor,
2207                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2208                                 cursor->trans->tid,
2209                                 cursor->trans->time32,
2210                                 0, NULL);
2211         }
2212         return(error);
2213 }
2214
2215 /*
2216  * Used to write a generic record w/optional data to the media b-tree
2217  * when no inode context is available.  Used by the mirroring and
2218  * snapshot code.
2219  *
2220  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2221  * flagged for backend operation and not flagged ASOF (since we are
2222  * doing an insertion).
2223  *
2224  * This function will acquire the appropriate sync lock and will set
2225  * the cursor insertion flag for the operation, do the btree lookup,
2226  * and the insertion, and clear the insertion flag and sync lock before
2227  * returning.  The cursor state will be such that the caller can continue
2228  * scanning (used by the mirroring code).
2229  *
2230  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2231  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2232  *
2233  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2234  *                retry.
2235  *
2236  *       EALREADY can be returned if the record already exists (WARNING,
2237  *                because ASOF cannot be used no check is made for illegal
2238  *                duplicates).
2239  *
2240  * NOTE: Do not use the function for normal inode-related records as this
2241  *       functions goes directly to the media and is not integrated with
2242  *       in-memory records.
2243  */
2244 int
2245 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2246                         void *udata, int mode)
2247 {
2248         hammer_transaction_t trans;
2249         hammer_buffer_t data_buffer;
2250         hammer_off_t ndata_offset;
2251         hammer_tid_t high_tid;
2252         void *ndata;
2253         int error;
2254         int doprop;
2255
2256         trans = cursor->trans;
2257         data_buffer = NULL;
2258         ndata_offset = 0;
2259         doprop = 0;
2260
2261         KKASSERT((cursor->flags &
2262                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2263                   (HAMMER_CURSOR_BACKEND));
2264
2265         hammer_sync_lock_sh(trans);
2266
2267         if (leaf->data_len) {
2268                 ndata = hammer_alloc_data(trans, leaf->data_len,
2269                                           leaf->base.rec_type,
2270                                           &ndata_offset, &data_buffer,
2271                                           0, &error);
2272                 if (ndata == NULL) {
2273                         hammer_sync_unlock(trans);
2274                         return (error);
2275                 }
2276                 leaf->data_offset = ndata_offset;
2277                 hammer_modify_buffer(trans, data_buffer, NULL, 0);
2278
2279                 switch(mode) {
2280                 case HAMMER_CREATE_MODE_UMIRROR:
2281                         error = copyin(udata, ndata, leaf->data_len);
2282                         if (error == 0) {
2283                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2284                                         kprintf("data crc mismatch on pipe\n");
2285                                         error = EINVAL;
2286                                 } else {
2287                                         error = hammer_cursor_localize_data(
2288                                                         ndata, leaf);
2289                                 }
2290                         }
2291                         break;
2292                 case HAMMER_CREATE_MODE_SYS:
2293                         bcopy(udata, ndata, leaf->data_len);
2294                         error = 0;
2295                         hammer_crc_set_leaf(ndata, leaf);
2296                         break;
2297                 default:
2298                         panic("hammer: hammer_create_at_cursor: bad mode %d",
2299                                 mode);
2300                         break; /* NOT REACHED */
2301                 }
2302                 hammer_modify_buffer_done(data_buffer);
2303         } else {
2304                 leaf->data_offset = 0;
2305                 error = 0;
2306                 ndata = NULL;
2307         }
2308         if (error)
2309                 goto failed;
2310
2311         /*
2312          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2313          */
2314         cursor->flags |= HAMMER_CURSOR_INSERT;
2315         error = hammer_btree_lookup(cursor);
2316         if (error != ENOENT) {
2317                 if (error == 0)
2318                         error = EALREADY;
2319                 goto failed;
2320         }
2321         error = hammer_btree_insert(cursor, leaf, &doprop);
2322
2323         /*
2324          * Cursor is left on current element, we want to skip it now.
2325          * (in case the caller is scanning)
2326          */
2327         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2328         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2329
2330         /*
2331          * If the insertion happens to be creating (and not just replacing)
2332          * an inode we have to track it.
2333          */
2334         if (error == 0 &&
2335             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2336             leaf->base.delete_tid == 0) {
2337                 hammer_modify_volume_field(trans, trans->rootvol,
2338                                            vol0_stat_inodes);
2339                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2340                 hammer_modify_volume_done(trans->rootvol);
2341         }
2342
2343         /*
2344          * vol0_next_tid must track the highest TID stored in the filesystem.
2345          * We do not need to generate undo for this update.
2346          */
2347         high_tid = leaf->base.create_tid;
2348         if (high_tid < leaf->base.delete_tid)
2349                 high_tid = leaf->base.delete_tid;
2350         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2351                 hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2352                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2353                 hammer_modify_volume_done(trans->rootvol);
2354         }
2355
2356         /*
2357          * WARNING!  cursor's leaf pointer may have changed after
2358          *           do_propagation returns.
2359          */
2360         if (error == 0 && doprop)
2361                 hammer_btree_do_propagation(cursor, NULL, leaf);
2362
2363 failed:
2364         /*
2365          * Cleanup
2366          */
2367         if (error && leaf->data_offset) {
2368                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2369
2370         }
2371         hammer_sync_unlock(trans);
2372         if (data_buffer)
2373                 hammer_rel_buffer(data_buffer, 0);
2374         return (error);
2375 }
2376
2377 /*
2378  * Delete the B-Tree element at the current cursor and do any necessary
2379  * mirror propagation.
2380  *
2381  * The cursor must be properly positioned for an iteration on return but
2382  * may be pointing at an internal element.
2383  *
2384  * An element can be un-deleted by passing a delete_tid of 0 with
2385  * HAMMER_DELETE_ADJUST.
2386  */
2387 int
2388 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2389                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2390                         int track, int64_t *stat_bytes)
2391 {
2392         struct hammer_btree_leaf_elm save_leaf;
2393         hammer_transaction_t trans;
2394         hammer_btree_leaf_elm_t leaf;
2395         hammer_node_t node;
2396         hammer_btree_elm_t elm;
2397         hammer_off_t data_offset;
2398         int32_t data_len;
2399         u_int16_t rec_type;
2400         int error;
2401         int icount;
2402         int doprop;
2403
2404         error = hammer_cursor_upgrade(cursor);
2405         if (error)
2406                 return(error);
2407
2408         trans = cursor->trans;
2409         node = cursor->node;
2410         elm = &node->ondisk->elms[cursor->index];
2411         leaf = &elm->leaf;
2412         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2413
2414         hammer_sync_lock_sh(trans);
2415         doprop = 0;
2416         icount = 0;
2417
2418         /*
2419          * Adjust the delete_tid.  Update the mirror_tid propagation field
2420          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2421          */
2422         if (delete_flags & HAMMER_DELETE_ADJUST) {
2423                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2424                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2425                                 icount = -1;
2426                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2427                                 icount = 1;
2428                 }
2429
2430                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2431                 elm->leaf.base.delete_tid = delete_tid;
2432                 elm->leaf.delete_ts = delete_ts;
2433                 hammer_modify_node_done(node);
2434
2435                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2436                         hammer_modify_node_field(trans, node, mirror_tid);
2437                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2438                         hammer_modify_node_done(node);
2439                         doprop = 1;
2440                         if (hammer_debug_general & 0x0002) {
2441                                 kprintf("delete_at_cursor: propagate %016llx"
2442                                         " @%016llx\n",
2443                                         (long long)elm->leaf.base.delete_tid,
2444                                         (long long)node->node_offset);
2445                         }
2446                 }
2447
2448                 /*
2449                  * Adjust for the iteration.  We have deleted the current
2450                  * element and want to clear ATEDISK so the iteration does
2451                  * not skip the element after, which now becomes the current
2452                  * element.  This element must be re-tested if doing an
2453                  * iteration, which is handled by the RETEST flag.
2454                  */
2455                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2456                         cursor->flags |= HAMMER_CURSOR_RETEST;
2457                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2458                 }
2459
2460                 /*
2461                  * An on-disk record cannot have the same delete_tid
2462                  * as its create_tid.  In a chain of record updates
2463                  * this could result in a duplicate record.
2464                  */
2465                 KKASSERT(elm->leaf.base.delete_tid !=
2466                          elm->leaf.base.create_tid);
2467         }
2468
2469         /*
2470          * Destroy the B-Tree element if asked (typically if a nohistory
2471          * file or mount, or when called by the pruning code).
2472          *
2473          * Adjust the ATEDISK flag to properly support iterations.
2474          */
2475         if (delete_flags & HAMMER_DELETE_DESTROY) {
2476                 data_offset = elm->leaf.data_offset;
2477                 data_len = elm->leaf.data_len;
2478                 rec_type = elm->leaf.base.rec_type;
2479                 if (doprop) {
2480                         save_leaf = elm->leaf;
2481                         leaf = &save_leaf;
2482                 }
2483                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2484                     elm->leaf.base.delete_tid == 0) {
2485                         icount = -1;
2486                 }
2487
2488                 error = hammer_btree_delete(cursor);
2489                 if (error == 0) {
2490                         /*
2491                          * The deletion moves the next element (if any) to
2492                          * the current element position.  We must clear
2493                          * ATEDISK so this element is not skipped and we
2494                          * must set RETEST to force any iteration to re-test
2495                          * the element.
2496                          */
2497                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2498                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2499                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2500                         }
2501                 }
2502                 if (error == 0) {
2503                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2504                         case HAMMER_ZONE_LARGE_DATA:
2505                         case HAMMER_ZONE_SMALL_DATA:
2506                         case HAMMER_ZONE_META:
2507                                 hammer_blockmap_free(trans,
2508                                                      data_offset, data_len);
2509                                 break;
2510                         default:
2511                                 break;
2512                         }
2513                 }
2514         }
2515
2516         /*
2517          * Track inode count and next_tid.  This is used by the mirroring
2518          * and PFS code.  icount can be negative, zero, or positive.
2519          */
2520         if (error == 0 && track) {
2521                 if (icount) {
2522                         hammer_modify_volume_field(trans, trans->rootvol,
2523                                                    vol0_stat_inodes);
2524                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2525                         hammer_modify_volume_done(trans->rootvol);
2526                 }
2527                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2528                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2529                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2530                         hammer_modify_volume_done(trans->rootvol);
2531                 }
2532         }
2533
2534         /*
2535          * mirror_tid propagation occurs if the node's mirror_tid had to be
2536          * updated while adjusting the delete_tid.
2537          *
2538          * This occurs when deleting even in nohistory mode, but does not
2539          * occur when pruning an already-deleted node.
2540          *
2541          * cursor->ip is NULL when called from the pruning, mirroring,
2542          * and pfs code.  If non-NULL propagation will be conditionalized
2543          * on whether the PFS is in no-history mode or not.
2544          *
2545          * WARNING: cursor's leaf pointer may have changed after do_propagation
2546          *          returns!
2547          */
2548         if (doprop) {
2549                 if (cursor->ip)
2550                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2551                 else
2552                         hammer_btree_do_propagation(cursor, NULL, leaf);
2553         }
2554         hammer_sync_unlock(trans);
2555         return (error);
2556 }
2557
2558 /*
2559  * Determine whether we can remove a directory.  This routine checks whether
2560  * a directory is empty or not and enforces flush connectivity.
2561  *
2562  * Flush connectivity requires that we block if the target directory is
2563  * currently flushing, otherwise it may not end up in the same flush group.
2564  *
2565  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2566  */
2567 int
2568 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2569 {
2570         struct hammer_cursor cursor;
2571         int error;
2572
2573         /*
2574          * Check directory empty
2575          */
2576         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2577
2578         cursor.key_beg.localization = ip->obj_localization +
2579                                       hammer_dir_localization(ip);
2580         cursor.key_beg.obj_id = ip->obj_id;
2581         cursor.key_beg.create_tid = 0;
2582         cursor.key_beg.delete_tid = 0;
2583         cursor.key_beg.obj_type = 0;
2584         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2585         cursor.key_beg.key = HAMMER_MIN_KEY;
2586
2587         cursor.key_end = cursor.key_beg;
2588         cursor.key_end.rec_type = 0xFFFF;
2589         cursor.key_end.key = HAMMER_MAX_KEY;
2590
2591         cursor.asof = ip->obj_asof;
2592         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2593
2594         error = hammer_ip_first(&cursor);
2595         if (error == ENOENT)
2596                 error = 0;
2597         else if (error == 0)
2598                 error = ENOTEMPTY;
2599         hammer_done_cursor(&cursor);
2600         return(error);
2601 }
2602
2603 /*
2604  * Localize the data payload.  Directory entries may need their
2605  * localization adjusted.
2606  */
2607 static
2608 int
2609 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2610                             hammer_btree_leaf_elm_t leaf)
2611 {
2612         u_int32_t localization;
2613
2614         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2615                 localization = leaf->base.localization &
2616                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2617                 if (data->entry.localization != localization) {
2618                         data->entry.localization = localization;
2619                         hammer_crc_set_leaf(data, leaf);
2620                 }
2621         }
2622         return(0);
2623 }