sys/vfs/hammer/hammer_object.c

   1 /*
   2  * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  *
  34  * $DragonFly: src/sys/vfs/hammer/hammer_object.c,v 1.97 2008/09/23 22:28:56 dillon Exp $
  35  */
  36
  37 #include "hammer.h"
  38
  39 static int hammer_mem_lookup(hammer_cursor_t cursor);
  40 static void hammer_mem_first(hammer_cursor_t cursor);
  41 static int hammer_frontend_trunc_callback(hammer_record_t record,
  42                                 void *data __unused);
  43 static int hammer_bulk_scan_callback(hammer_record_t record, void *data);
  44 static int hammer_record_needs_overwrite_delete(hammer_record_t record);
  45 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
  46                                 hammer_btree_leaf_elm_t leaf);
  47 static int hammer_cursor_localize_data(hammer_data_ondisk_t data,
  48                                 hammer_btree_leaf_elm_t leaf);
  49
  50 struct rec_trunc_info {
  51         u_int16_t       rec_type;
  52         int64_t         trunc_off;
  53 };
  54
  55 struct hammer_bulk_info {
  56         hammer_record_t record;
  57         hammer_record_t conflict;
  58 };
  59
  60 /*
  61  * Red-black tree support.  Comparison code for insertion.
  62  */
  63 static int
  64 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2)
  65 {
  66         if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type)
  67                 return(-1);
  68         if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type)
  69                 return(1);
  70
  71         if (rec1->leaf.base.key < rec2->leaf.base.key)
  72                 return(-1);
  73         if (rec1->leaf.base.key > rec2->leaf.base.key)
  74                 return(1);
  75
  76         /*
  77          * For search & insertion purposes records deleted by the
  78          * frontend or deleted/committed by the backend are silently
  79          * ignored.  Otherwise pipelined insertions will get messed
  80          * up.
  81          *
  82          * rec1 is greater then rec2 if rec1 is marked deleted.
  83          * rec1 is less then rec2 if rec2 is marked deleted.
  84          *
  85          * Multiple deleted records may be present, do not return 0
  86          * if both are marked deleted.
  87          */
  88         if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  89                            HAMMER_RECF_COMMITTED)) {
  90                 return(1);
  91         }
  92         if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
  93                            HAMMER_RECF_COMMITTED)) {
  94                 return(-1);
  95         }
  96
  97         return(0);
  98 }
  99
 100 /*
 101  * Basic record comparison code similar to hammer_btree_cmp().
 102  *
 103  * obj_id is not compared and may not yet be assigned in the record.
 104  */
 105 static int
 106 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec)
 107 {
 108         if (elm->rec_type < rec->leaf.base.rec_type)
 109                 return(-3);
 110         if (elm->rec_type > rec->leaf.base.rec_type)
 111                 return(3);
 112
 113         if (elm->key < rec->leaf.base.key)
 114                 return(-2);
 115         if (elm->key > rec->leaf.base.key)
 116                 return(2);
 117
 118         /*
 119          * Never match against an item deleted by the frontend
 120          * or backend, or committed by the backend.
 121          *
 122          * elm is less then rec if rec is marked deleted.
 123          */
 124         if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 125                           HAMMER_RECF_COMMITTED)) {
 126                 return(-1);
 127         }
 128         return(0);
 129 }
 130
 131 /*
 132  * Ranged scan to locate overlapping record(s).  This is used by
 133  * hammer_ip_get_bulk() to locate an overlapping record.  We have
 134  * to use a ranged scan because the keys for data records with the
 135  * same file base offset can be different due to differing data_len's.
 136  *
 137  * NOTE: The base file offset of a data record is (key - data_len), not (key).
 138  */
 139 static int
 140 hammer_rec_overlap_cmp(hammer_record_t rec, void *data)
 141 {
 142         struct hammer_bulk_info *info = data;
 143         hammer_btree_leaf_elm_t leaf = &info->record->leaf;
 144
 145         if (rec->leaf.base.rec_type < leaf->base.rec_type)
 146                 return(-3);
 147         if (rec->leaf.base.rec_type > leaf->base.rec_type)
 148                 return(3);
 149
 150         /*
 151          * Overlap compare
 152          */
 153         if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
 154                 /* rec_beg >= leaf_end */
 155                 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key)
 156                         return(2);
 157                 /* rec_end <= leaf_beg */
 158                 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len)
 159                         return(-2);
 160         } else {
 161                 if (rec->leaf.base.key < leaf->base.key)
 162                         return(-2);
 163                 if (rec->leaf.base.key > leaf->base.key)
 164                         return(2);
 165         }
 166
 167         /*
 168          * We have to return 0 at this point, even if DELETED_FE is set,
 169          * because returning anything else will cause the scan to ignore
 170          * one of the branches when we really want it to check both.
 171          */
 172         return(0);
 173 }
 174
 175 /*
 176  * RB_SCAN comparison code for hammer_mem_first().  The argument order
 177  * is reversed so the comparison result has to be negated.  key_beg and
 178  * key_end are both range-inclusive.
 179  *
 180  * Localized deletions are not cached in-memory.
 181  */
 182 static
 183 int
 184 hammer_rec_scan_cmp(hammer_record_t rec, void *data)
 185 {
 186         hammer_cursor_t cursor = data;
 187         int r;
 188
 189         r = hammer_rec_cmp(&cursor->key_beg, rec);
 190         if (r > 1)
 191                 return(-1);
 192         r = hammer_rec_cmp(&cursor->key_end, rec);
 193         if (r < -1)
 194                 return(1);
 195         return(0);
 196 }
 197
 198 /*
 199  * This compare function is used when simply looking up key_beg.
 200  */
 201 static
 202 int
 203 hammer_rec_find_cmp(hammer_record_t rec, void *data)
 204 {
 205         hammer_cursor_t cursor = data;
 206         int r;
 207
 208         r = hammer_rec_cmp(&cursor->key_beg, rec);
 209         if (r > 1)
 210                 return(-1);
 211         if (r < -1)
 212                 return(1);
 213         return(0);
 214 }
 215
 216 /*
 217  * Locate blocks within the truncation range.  Partial blocks do not count.
 218  */
 219 static
 220 int
 221 hammer_rec_trunc_cmp(hammer_record_t rec, void *data)
 222 {
 223         struct rec_trunc_info *info = data;
 224
 225         if (rec->leaf.base.rec_type < info->rec_type)
 226                 return(-1);
 227         if (rec->leaf.base.rec_type > info->rec_type)
 228                 return(1);
 229
 230         switch(rec->leaf.base.rec_type) {
 231         case HAMMER_RECTYPE_DB:
 232                 /*
 233                  * DB record key is not beyond the truncation point, retain.
 234                  */
 235                 if (rec->leaf.base.key < info->trunc_off)
 236                         return(-1);
 237                 break;
 238         case HAMMER_RECTYPE_DATA:
 239                 /*
 240                  * DATA record offset start is not beyond the truncation point,
 241                  * retain.
 242                  */
 243                 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off)
 244                         return(-1);
 245                 break;
 246         default:
 247                 panic("hammer_rec_trunc_cmp: unexpected record type");
 248         }
 249
 250         /*
 251          * The record start is >= the truncation point, return match,
 252          * the record should be destroyed.
 253          */
 254         return(0);
 255 }
 256
 257 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare);
 258
 259 /*
 260  * Allocate a record for the caller to finish filling in.  The record is
 261  * returned referenced.
 262  */
 263 hammer_record_t
 264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len)
 265 {
 266         hammer_record_t record;
 267         hammer_mount_t hmp;
 268
 269         hmp = ip->hmp;
 270         ++hammer_count_records;
 271         record = kmalloc(sizeof(*record), hmp->m_misc,
 272                          M_WAITOK | M_ZERO | M_USE_RESERVE);
 273         record->flush_state = HAMMER_FST_IDLE;
 274         record->ip = ip;
 275         record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
 276         record->leaf.data_len = data_len;
 277         hammer_ref(&record->lock);
 278
 279         if (data_len) {
 280                 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO);
 281                 record->flags |= HAMMER_RECF_ALLOCDATA;
 282                 ++hammer_count_record_datas;
 283         }
 284
 285         return (record);
 286 }
 287
 288 void
 289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident)
 290 {
 291         while (record->flush_state == HAMMER_FST_FLUSH) {
 292                 record->flags |= HAMMER_RECF_WANTED;
 293                 tsleep(record, 0, ident, 0);
 294         }
 295 }
 296
 297 /*
 298  * Called from the backend, hammer_inode.c, after a record has been
 299  * flushed to disk.  The record has been exclusively locked by the
 300  * caller and interlocked with BE.
 301  *
 302  * We clean up the state, unlock, and release the record (the record
 303  * was referenced by the fact that it was in the HAMMER_FST_FLUSH state).
 304  */
 305 void
 306 hammer_flush_record_done(hammer_record_t record, int error)
 307 {
 308         hammer_inode_t target_ip;
 309
 310         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
 311         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
 312
 313         /*
 314          * If an error occured, the backend was unable to sync the
 315          * record to its media.  Leave the record intact.
 316          */
 317         if (error) {
 318                 hammer_critical_error(record->ip->hmp, record->ip, error,
 319                                       "while flushing record");
 320         }
 321
 322         --record->flush_group->refs;
 323         record->flush_group = NULL;
 324
 325         /*
 326          * Adjust the flush state and dependancy based on success or
 327          * failure.
 328          */
 329         if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
 330                 if ((target_ip = record->target_ip) != NULL) {
 331                         TAILQ_REMOVE(&target_ip->target_list, record,
 332                                      target_entry);
 333                         record->target_ip = NULL;
 334                         hammer_test_inode(target_ip);
 335                 }
 336                 record->flush_state = HAMMER_FST_IDLE;
 337         } else {
 338                 if (record->target_ip) {
 339                         record->flush_state = HAMMER_FST_SETUP;
 340                         hammer_test_inode(record->ip);
 341                         hammer_test_inode(record->target_ip);
 342                 } else {
 343                         record->flush_state = HAMMER_FST_IDLE;
 344                 }
 345         }
 346         record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
 347
 348         /*
 349          * Cleanup
 350          */
 351         if (record->flags & HAMMER_RECF_WANTED) {
 352                 record->flags &= ~HAMMER_RECF_WANTED;
 353                 wakeup(record);
 354         }
 355         hammer_rel_mem_record(record);
 356 }
 357
 358 /*
 359  * Release a memory record.  Records marked for deletion are immediately
 360  * removed from the RB-Tree but otherwise left intact until the last ref
 361  * goes away.
 362  */
 363 void
 364 hammer_rel_mem_record(struct hammer_record *record)
 365 {
 366         hammer_mount_t hmp;
 367         hammer_reserve_t resv;
 368         hammer_inode_t ip;
 369         hammer_inode_t target_ip;
 370         int diddrop;
 371
 372         hammer_rel(&record->lock);
 373
 374         if (hammer_norefs(&record->lock)) {
 375                 /*
 376                  * Upon release of the last reference wakeup any waiters.
 377                  * The record structure may get destroyed so callers will
 378                  * loop up and do a relookup.
 379                  *
 380                  * WARNING!  Record must be removed from RB-TREE before we
 381                  * might possibly block.  hammer_test_inode() can block!
 382                  */
 383                 ip = record->ip;
 384                 hmp = ip->hmp;
 385
 386                 /*
 387                  * Upon release of the last reference a record marked deleted
 388                  * by the front or backend, or committed by the backend,
 389                  * is destroyed.
 390                  */
 391                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 392                                      HAMMER_RECF_DELETED_BE |
 393                                      HAMMER_RECF_COMMITTED)) {
 394                         KKASSERT(hammer_isactive(&ip->lock) > 0);
 395                         KKASSERT(record->flush_state != HAMMER_FST_FLUSH);
 396
 397                         /*
 398                          * target_ip may have zero refs, we have to ref it
 399                          * to prevent it from being ripped out from under
 400                          * us.
 401                          */
 402                         if ((target_ip = record->target_ip) != NULL) {
 403                                 TAILQ_REMOVE(&target_ip->target_list,
 404                                              record, target_entry);
 405                                 record->target_ip = NULL;
 406                                 hammer_ref(&target_ip->lock);
 407                         }
 408
 409                         /*
 410                          * Remove the record from the B-Tree
 411                          */
 412                         if (record->flags & HAMMER_RECF_ONRBTREE) {
 413                                 RB_REMOVE(hammer_rec_rb_tree,
 414                                           &record->ip->rec_tree,
 415                                           record);
 416                                 record->flags &= ~HAMMER_RECF_ONRBTREE;
 417                                 KKASSERT(ip->rsv_recs > 0);
 418                                 if (RB_EMPTY(&record->ip->rec_tree)) {
 419                                         record->ip->flags &=
 420                                                         ~HAMMER_INODE_XDIRTY;
 421                                         record->ip->sync_flags &=
 422                                                         ~HAMMER_INODE_XDIRTY;
 423                                 }
 424                                 diddrop = 1;
 425                         } else {
 426                                 diddrop = 0;
 427                         }
 428
 429                         /*
 430                          * We must wait for any direct-IO to complete before
 431                          * we can destroy the record because the bio may
 432                          * have a reference to it.
 433                          */
 434                         if (record->gflags &
 435                            (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) {
 436                                 hammer_io_direct_wait(record);
 437                         }
 438
 439                         /*
 440                          * Account for the completion after the direct IO
 441                          * has completed.
 442                          */
 443                         if (diddrop) {
 444                                 --hmp->rsv_recs;
 445                                 --ip->rsv_recs;
 446                                 hmp->rsv_databytes -= record->leaf.data_len;
 447
 448                                 if (RB_EMPTY(&record->ip->rec_tree))
 449                                         hammer_test_inode(record->ip);
 450                                 if (ip->rsv_recs == hammer_limit_inode_recs - 1)
 451                                         wakeup(&ip->rsv_recs);
 452                         }
 453
 454                         /*
 455                          * Do this test after removing record from the B-Tree.
 456                          */
 457                         if (target_ip) {
 458                                 hammer_test_inode(target_ip);
 459                                 hammer_rel_inode(target_ip, 0);
 460                         }
 461
 462                         if (record->flags & HAMMER_RECF_ALLOCDATA) {
 463                                 --hammer_count_record_datas;
 464                                 kfree(record->data, hmp->m_misc);
 465                                 record->flags &= ~HAMMER_RECF_ALLOCDATA;
 466                         }
 467
 468                         /*
 469                          * Release the reservation.
 470                          *
 471                          * If the record was not committed we can theoretically
 472                          * undo the reservation.  However, doing so might
 473                          * create weird edge cases with the ordering of
 474                          * direct writes because the related buffer cache
 475                          * elements are per-vnode.  So we don't try.
 476                          */
 477                         if ((resv = record->resv) != NULL) {
 478                                 /* XXX undo leaf.data_offset,leaf.data_len */
 479                                 hammer_blockmap_reserve_complete(hmp, resv);
 480                                 record->resv = NULL;
 481                         }
 482                         record->data = NULL;
 483                         --hammer_count_records;
 484                         kfree(record, hmp->m_misc);
 485                 }
 486         }
 487 }
 488
 489 /*
 490  * Record visibility depends on whether the record is being accessed by
 491  * the backend or the frontend.  Backend tests ignore the frontend delete
 492  * flag.  Frontend tests do NOT ignore the backend delete/commit flags and
 493  * must also check for commit races.
 494  *
 495  * Return non-zero if the record is visible, zero if it isn't or if it is
 496  * deleted.  Returns 0 if the record has been comitted (unless the special
 497  * delete-visibility flag is set).  A committed record must be located
 498  * via the media B-Tree.  Returns non-zero if the record is good.
 499  *
 500  * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory
 501  * records to be returned.  This is so pending deletions are detected
 502  * when using an iterator to locate an unused hash key, or when we need
 503  * to locate historical records on-disk to destroy.
 504  */
 505 static __inline
 506 int
 507 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record)
 508 {
 509         if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY)
 510                 return(1);
 511         if (cursor->flags & HAMMER_CURSOR_BACKEND) {
 512                 if (record->flags & (HAMMER_RECF_DELETED_BE |
 513                                      HAMMER_RECF_COMMITTED)) {
 514                         return(0);
 515                 }
 516         } else {
 517                 if (record->flags & (HAMMER_RECF_DELETED_FE |
 518                                      HAMMER_RECF_DELETED_BE |
 519                                      HAMMER_RECF_COMMITTED)) {
 520                         return(0);
 521                 }
 522         }
 523         return(1);
 524 }
 525
 526 /*
 527  * This callback is used as part of the RB_SCAN function for in-memory
 528  * records.  We terminate it (return -1) as soon as we get a match.
 529  *
 530  * This routine is used by frontend code.
 531  *
 532  * The primary compare code does not account for ASOF lookups.  This
 533  * code handles that case as well as a few others.
 534  */
 535 static
 536 int
 537 hammer_rec_scan_callback(hammer_record_t rec, void *data)
 538 {
 539         hammer_cursor_t cursor = data;
 540
 541         /*
 542          * We terminate on success, so this should be NULL on entry.
 543          */
 544         KKASSERT(cursor->iprec == NULL);
 545
 546         /*
 547          * Skip if the record was marked deleted or committed.
 548          */
 549         if (hammer_ip_iterate_mem_good(cursor, rec) == 0)
 550                 return(0);
 551
 552         /*
 553          * Skip if not visible due to our as-of TID
 554          */
 555         if (cursor->flags & HAMMER_CURSOR_ASOF) {
 556                 if (cursor->asof < rec->leaf.base.create_tid)
 557                         return(0);
 558                 if (rec->leaf.base.delete_tid &&
 559                     cursor->asof >= rec->leaf.base.delete_tid) {
 560                         return(0);
 561                 }
 562         }
 563
 564         /*
 565          * ref the record.  The record is protected from backend B-Tree
 566          * interactions by virtue of the cursor's IP lock.
 567          */
 568         hammer_ref(&rec->lock);
 569
 570         /*
 571          * The record may have been deleted or committed while we
 572          * were blocked.  XXX remove?
 573          */
 574         if (hammer_ip_iterate_mem_good(cursor, rec) == 0) {
 575                 hammer_rel_mem_record(rec);
 576                 return(0);
 577         }
 578
 579         /*
 580          * Set the matching record and stop the scan.
 581          */
 582         cursor->iprec = rec;
 583         return(-1);
 584 }
 585
 586
 587 /*
 588  * Lookup an in-memory record given the key specified in the cursor.  Works
 589  * just like hammer_btree_lookup() but operates on an inode's in-memory
 590  * record list.
 591  *
 592  * The lookup must fail if the record is marked for deferred deletion.
 593  *
 594  * The API for mem/btree_lookup() does not mess with the ATE/EOF bits.
 595  */
 596 static
 597 int
 598 hammer_mem_lookup(hammer_cursor_t cursor)
 599 {
 600         KKASSERT(cursor->ip);
 601         if (cursor->iprec) {
 602                 hammer_rel_mem_record(cursor->iprec);
 603                 cursor->iprec = NULL;
 604         }
 605         hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp,
 606                                    hammer_rec_scan_callback, cursor);
 607
 608         return (cursor->iprec ? 0 : ENOENT);
 609 }
 610
 611 /*
 612  * hammer_mem_first() - locate the first in-memory record matching the
 613  * cursor within the bounds of the key range.
 614  *
 615  * WARNING!  API is slightly different from btree_first().  hammer_mem_first()
 616  * will set ATEMEM the same as MEMEOF, and does not return any error.
 617  */
 618 static
 619 void
 620 hammer_mem_first(hammer_cursor_t cursor)
 621 {
 622         hammer_inode_t ip;
 623
 624         ip = cursor->ip;
 625         KKASSERT(ip != NULL);
 626
 627         if (cursor->iprec) {
 628                 hammer_rel_mem_record(cursor->iprec);
 629                 cursor->iprec = NULL;
 630         }
 631         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_scan_cmp,
 632                                    hammer_rec_scan_callback, cursor);
 633
 634         if (cursor->iprec)
 635                 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM);
 636         else
 637                 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM;
 638 }
 639
 640 /************************************************************************
 641  *                   HAMMER IN-MEMORY RECORD FUNCTIONS                  *
 642  ************************************************************************
 643  *
 644  * These functions manipulate in-memory records.  Such records typically
 645  * exist prior to being committed to disk or indexed via the on-disk B-Tree.
 646  */
 647
 648 /*
 649  * Add a directory entry (dip,ncp) which references inode (ip).
 650  *
 651  * Note that the low 32 bits of the namekey are set temporarily to create
 652  * a unique in-memory record, and may be modified a second time when the
 653  * record is synchronized to disk.  In particular, the low 32 bits cannot be
 654  * all 0's when synching to disk, which is not handled here.
 655  *
 656  * NOTE: bytes does not include any terminating \0 on name, and name might
 657  * not be terminated.
 658  */
 659 int
 660 hammer_ip_add_directory(struct hammer_transaction *trans,
 661                      struct hammer_inode *dip, const char *name, int bytes,
 662                      struct hammer_inode *ip)
 663 {
 664         struct hammer_cursor cursor;
 665         hammer_record_t record;
 666         int error;
 667         u_int32_t max_iterations;
 668
 669         record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes));
 670
 671         record->type = HAMMER_MEM_RECORD_ADD;
 672         record->leaf.base.localization = dip->obj_localization +
 673                                          hammer_dir_localization(dip);
 674         record->leaf.base.obj_id = dip->obj_id;
 675         record->leaf.base.key = hammer_directory_namekey(dip, name, bytes,
 676                                                          &max_iterations);
 677         record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY;
 678         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 679         record->data->entry.obj_id = ip->obj_id;
 680         record->data->entry.localization = ip->obj_localization;
 681         bcopy(name, record->data->entry.name, bytes);
 682
 683         ++ip->ino_data.nlinks;
 684         ip->ino_data.ctime = trans->time;
 685         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 686
 687         /*
 688          * Find an unused namekey.  Both the in-memory record tree and
 689          * the B-Tree are checked.  We do not want historically deleted
 690          * names to create a collision as our iteration space may be limited,
 691          * and since create_tid wouldn't match anyway an ASOF search
 692          * must be used to locate collisions.
 693          *
 694          * delete-visibility is set so pending deletions do not give us
 695          * a false-negative on our ability to use an iterator.
 696          *
 697          * The iterator must not rollover the key.  Directory keys only
 698          * use the positive key space.
 699          */
 700         hammer_init_cursor(trans, &cursor, &dip->cache[1], dip);
 701         cursor.key_beg = record->leaf.base;
 702         cursor.flags |= HAMMER_CURSOR_ASOF;
 703         cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
 704         cursor.asof = ip->obj_asof;
 705
 706         while (hammer_ip_lookup(&cursor) == 0) {
 707                 ++record->leaf.base.key;
 708                 KKASSERT(record->leaf.base.key > 0);
 709                 cursor.key_beg.key = record->leaf.base.key;
 710                 if (--max_iterations == 0) {
 711                         hammer_rel_mem_record(record);
 712                         error = ENOSPC;
 713                         goto failed;
 714                 }
 715         }
 716
 717         /*
 718          * The target inode and the directory entry are bound together.
 719          */
 720         record->target_ip = ip;
 721         record->flush_state = HAMMER_FST_SETUP;
 722         TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry);
 723
 724         /*
 725          * The inode now has a dependancy and must be taken out of the idle
 726          * state.  An inode not in an idle state is given an extra reference.
 727          *
 728          * When transitioning to a SETUP state flag for an automatic reflush
 729          * when the dependancies are disposed of if someone is waiting on
 730          * the inode.
 731          */
 732         if (ip->flush_state == HAMMER_FST_IDLE) {
 733                 hammer_ref(&ip->lock);
 734                 ip->flush_state = HAMMER_FST_SETUP;
 735                 if (ip->flags & HAMMER_INODE_FLUSHW)
 736                         ip->flags |= HAMMER_INODE_REFLUSH;
 737         }
 738         error = hammer_mem_add(record);
 739         if (error == 0) {
 740                 dip->ino_data.mtime = trans->time;
 741                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 742         }
 743 failed:
 744         hammer_done_cursor(&cursor);
 745         return(error);
 746 }
 747
 748 /*
 749  * Delete the directory entry and update the inode link count.  The
 750  * cursor must be seeked to the directory entry record being deleted.
 751  *
 752  * The related inode should be share-locked by the caller.  The caller is
 753  * on the frontend.  It could also be NULL indicating that the directory
 754  * entry being removed has no related inode.
 755  *
 756  * This function can return EDEADLK requiring the caller to terminate
 757  * the cursor, any locks, wait on the returned record, and retry.
 758  */
 759 int
 760 hammer_ip_del_directory(struct hammer_transaction *trans,
 761                      hammer_cursor_t cursor, struct hammer_inode *dip,
 762                      struct hammer_inode *ip)
 763 {
 764         hammer_record_t record;
 765         int error;
 766
 767         if (hammer_cursor_inmem(cursor)) {
 768                 /*
 769                  * In-memory (unsynchronized) records can simply be freed.
 770                  *
 771                  * Even though the HAMMER_RECF_DELETED_FE flag is ignored
 772                  * by the backend, we must still avoid races against the
 773                  * backend potentially syncing the record to the media.
 774                  *
 775                  * We cannot call hammer_ip_delete_record(), that routine may
 776                  * only be called from the backend.
 777                  */
 778                 record = cursor->iprec;
 779                 if (record->flags & (HAMMER_RECF_INTERLOCK_BE |
 780                                      HAMMER_RECF_DELETED_BE |
 781                                      HAMMER_RECF_COMMITTED)) {
 782                         KKASSERT(cursor->deadlk_rec == NULL);
 783                         hammer_ref(&record->lock);
 784                         cursor->deadlk_rec = record;
 785                         error = EDEADLK;
 786                 } else {
 787                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
 788                         record->flags |= HAMMER_RECF_DELETED_FE;
 789                         error = 0;
 790                 }
 791         } else {
 792                 /*
 793                  * If the record is on-disk we have to queue the deletion by
 794                  * the record's key.  This also causes lookups to skip the
 795                  * record (lookups for the purposes of finding an unused
 796                  * directory key do not skip the record).
 797                  */
 798                 KKASSERT(dip->flags &
 799                          (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK));
 800                 record = hammer_alloc_mem_record(dip, 0);
 801                 record->type = HAMMER_MEM_RECORD_DEL;
 802                 record->leaf.base = cursor->leaf->base;
 803                 KKASSERT(dip->obj_id == record->leaf.base.obj_id);
 804
 805                 /*
 806                  * ip may be NULL, indicating the deletion of a directory
 807                  * entry which has no related inode.
 808                  */
 809                 record->target_ip = ip;
 810                 if (ip) {
 811                         record->flush_state = HAMMER_FST_SETUP;
 812                         TAILQ_INSERT_TAIL(&ip->target_list, record,
 813                                           target_entry);
 814                 } else {
 815                         record->flush_state = HAMMER_FST_IDLE;
 816                 }
 817
 818                 /*
 819                  * The inode now has a dependancy and must be taken out of
 820                  * the idle state.  An inode not in an idle state is given
 821                  * an extra reference.
 822                  *
 823                  * When transitioning to a SETUP state flag for an automatic
 824                  * reflush when the dependancies are disposed of if someone
 825                  * is waiting on the inode.
 826                  */
 827                 if (ip && ip->flush_state == HAMMER_FST_IDLE) {
 828                         hammer_ref(&ip->lock);
 829                         ip->flush_state = HAMMER_FST_SETUP;
 830                         if (ip->flags & HAMMER_INODE_FLUSHW)
 831                                 ip->flags |= HAMMER_INODE_REFLUSH;
 832                 }
 833
 834                 error = hammer_mem_add(record);
 835         }
 836
 837         /*
 838          * One less link.  The file may still be open in the OS even after
 839          * all links have gone away.
 840          *
 841          * We have to terminate the cursor before syncing the inode to
 842          * avoid deadlocking against ourselves.  XXX this may no longer
 843          * be true.
 844          *
 845          * If nlinks drops to zero and the vnode is inactive (or there is
 846          * no vnode), call hammer_inode_unloadable_check() to zonk the
 847          * inode.  If we don't do this here the inode will not be destroyed
 848          * on-media until we unmount.
 849          */
 850         if (error == 0) {
 851                 if (ip) {
 852                         --ip->ino_data.nlinks;  /* do before we might block */
 853                         ip->ino_data.ctime = trans->time;
 854                 }
 855                 dip->ino_data.mtime = trans->time;
 856                 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME);
 857                 if (ip) {
 858                         hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
 859                         if (ip->ino_data.nlinks == 0 &&
 860                             (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) {
 861                                 hammer_done_cursor(cursor);
 862                                 hammer_inode_unloadable_check(ip, 1);
 863                                 hammer_flush_inode(ip, 0);
 864                         }
 865                 }
 866
 867         }
 868         return(error);
 869 }
 870
 871 /*
 872  * Add a record to an inode.
 873  *
 874  * The caller must allocate the record with hammer_alloc_mem_record(ip) and
 875  * initialize the following additional fields:
 876  *
 877  * The related inode should be share-locked by the caller.  The caller is
 878  * on the frontend.
 879  *
 880  * record->rec.entry.base.base.key
 881  * record->rec.entry.base.base.rec_type
 882  * record->rec.entry.base.base.data_len
 883  * record->data         (a copy will be kmalloc'd if it cannot be embedded)
 884  */
 885 int
 886 hammer_ip_add_record(struct hammer_transaction *trans, hammer_record_t record)
 887 {
 888         hammer_inode_t ip = record->ip;
 889         int error;
 890
 891         KKASSERT(record->leaf.base.localization != 0);
 892         record->leaf.base.obj_id = ip->obj_id;
 893         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 894         error = hammer_mem_add(record);
 895         return(error);
 896 }
 897
 898 /*
 899  * Locate a pre-existing bulk record in memory.  The caller wishes to
 900  * replace the record with a new one.  The existing record may have a
 901  * different length (and thus a different key) so we have to use an
 902  * overlap check function.
 903  */
 904 static hammer_record_t
 905 hammer_ip_get_bulk(hammer_record_t record)
 906 {
 907         struct hammer_bulk_info info;
 908         hammer_inode_t ip = record->ip;
 909
 910         info.record = record;
 911         info.conflict = NULL;
 912         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp,
 913                                    hammer_bulk_scan_callback, &info);
 914
 915         return(info.conflict);  /* may be NULL */
 916 }
 917
 918 /*
 919  * Take records vetted by overlap_cmp.  The first non-deleted record
 920  * (if any) stops the scan.
 921  */
 922 static int
 923 hammer_bulk_scan_callback(hammer_record_t record, void *data)
 924 {
 925         struct hammer_bulk_info *info = data;
 926
 927         if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
 928                              HAMMER_RECF_COMMITTED)) {
 929                 return(0);
 930         }
 931         hammer_ref(&record->lock);
 932         info->conflict = record;
 933         return(-1);                     /* stop scan */
 934 }
 935
 936 /*
 937  * Reserve blockmap space placemarked with an in-memory record.
 938  *
 939  * This routine is called by the frontend in order to be able to directly
 940  * flush a buffer cache buffer.  The frontend has locked the related buffer
 941  * cache buffers and we should be able to manipulate any overlapping
 942  * in-memory records.
 943  *
 944  * The caller is responsible for adding the returned record and deleting
 945  * the returned conflicting record (if any), typically by calling
 946  * hammer_ip_replace_bulk() (via hammer_io_direct_write()).
 947  */
 948 hammer_record_t
 949 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes,
 950                    int *errorp)
 951 {
 952         hammer_record_t record;
 953         int zone;
 954
 955         /*
 956          * Create a record to cover the direct write.  The record cannot
 957          * be added to the in-memory RB tree here as it might conflict
 958          * with an existing memory record.  See hammer_io_direct_write().
 959          *
 960          * The backend is responsible for finalizing the space reserved in
 961          * this record.
 962          *
 963          * XXX bytes not aligned, depend on the reservation code to
 964          * align the reservation.
 965          */
 966         record = hammer_alloc_mem_record(ip, 0);
 967         zone = (bytes >= HAMMER_BUFSIZE) ? HAMMER_ZONE_LARGE_DATA_INDEX :
 968                                            HAMMER_ZONE_SMALL_DATA_INDEX;
 969         record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes,
 970                                                &record->leaf.data_offset,
 971                                                errorp);
 972         if (record->resv == NULL) {
 973                 kprintf("hammer_ip_add_bulk: reservation failed\n");
 974                 hammer_rel_mem_record(record);
 975                 return(NULL);
 976         }
 977         record->type = HAMMER_MEM_RECORD_DATA;
 978         record->leaf.base.rec_type = HAMMER_RECTYPE_DATA;
 979         record->leaf.base.obj_type = ip->ino_leaf.base.obj_type;
 980         record->leaf.base.obj_id = ip->obj_id;
 981         record->leaf.base.key = file_offset + bytes;
 982         record->leaf.base.localization = ip->obj_localization +
 983                                          HAMMER_LOCALIZE_MISC;
 984         record->leaf.data_len = bytes;
 985         hammer_crc_set_leaf(data, &record->leaf);
 986         KKASSERT(*errorp == 0);
 987
 988         return(record);
 989 }
 990
 991 /*
 992  * Called by hammer_io_direct_write() prior to any possible completion
 993  * of the BIO to emplace the memory record associated with the I/O and
 994  * to replace any prior memory record which might still be active.
 995  *
 996  * Setting the FE deleted flag on the old record (if any) avoids any RB
 997  * tree insertion conflict, amoung other things.
 998  *
 999  * This has to be done prior to the caller completing any related buffer
1000  * cache I/O or a reinstantiation of the buffer may load data from the
1001  * old media location instead of the new media location.  The holding
1002  * of the locked buffer cache buffer serves to interlock the record
1003  * replacement operation.
1004  */
1005 void
1006 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record)
1007 {
1008         hammer_record_t conflict;
1009         int error;
1010
1011         while ((conflict = hammer_ip_get_bulk(record)) != NULL) {
1012                 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) {
1013                         conflict->flags |= HAMMER_RECF_DELETED_FE;
1014                         break;
1015                 }
1016                 conflict->flags |= HAMMER_RECF_WANTED;
1017                 tsleep(conflict, 0, "hmrrc3", 0);
1018                 hammer_rel_mem_record(conflict);
1019         }
1020         error = hammer_mem_add(record);
1021         if (conflict)
1022                 hammer_rel_mem_record(conflict);
1023         KKASSERT(error == 0);
1024 }
1025
1026 /*
1027  * Frontend truncation code.  Scan in-memory records only.  On-disk records
1028  * and records in a flushing state are handled by the backend.  The vnops
1029  * setattr code will handle the block containing the truncation point.
1030  *
1031  * Partial blocks are not deleted.
1032  *
1033  * This code is only called on regular files.
1034  */
1035 int
1036 hammer_ip_frontend_trunc(struct hammer_inode *ip, off_t file_size)
1037 {
1038         struct rec_trunc_info info;
1039
1040         switch(ip->ino_data.obj_type) {
1041         case HAMMER_OBJTYPE_REGFILE:
1042                 info.rec_type = HAMMER_RECTYPE_DATA;
1043                 break;
1044         case HAMMER_OBJTYPE_DBFILE:
1045                 info.rec_type = HAMMER_RECTYPE_DB;
1046                 break;
1047         default:
1048                 return(EINVAL);
1049         }
1050         info.trunc_off = file_size;
1051         hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp,
1052                                    hammer_frontend_trunc_callback, &info);
1053         return(0);
1054 }
1055
1056 /*
1057  * Scan callback for frontend records to destroy during a truncation.
1058  * We must ensure that DELETED_FE is set on the record or the frontend
1059  * will get confused in future read() calls.
1060  *
1061  * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held.
1062  *       In this rare case we must wait for the interlock to be cleared.
1063  *
1064  * NOTE: This function is only called on regular files.  There are further
1065  *       restrictions to the setting of DELETED_FE on directory records
1066  *       undergoing a flush due to sensitive inode link count calculations.
1067  */
1068 static int
1069 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused)
1070 {
1071         if (record->flags & HAMMER_RECF_DELETED_FE)
1072                 return(0);
1073 #if 0
1074         if (record->flush_state == HAMMER_FST_FLUSH)
1075                 return(0);
1076 #endif
1077         hammer_ref(&record->lock);
1078         while (record->flags & HAMMER_RECF_INTERLOCK_BE)
1079                 hammer_wait_mem_record_ident(record, "hmmtrr");
1080         record->flags |= HAMMER_RECF_DELETED_FE;
1081         hammer_rel_mem_record(record);
1082         return(0);
1083 }
1084
1085 /*
1086  * Return 1 if the caller must check for and delete existing records
1087  * before writing out a new data record.
1088  *
1089  * Return 0 if the caller can just insert the record into the B-Tree without
1090  * checking.
1091  */
1092 static int
1093 hammer_record_needs_overwrite_delete(hammer_record_t record)
1094 {
1095         hammer_inode_t ip = record->ip;
1096         int64_t file_offset;
1097         int r;
1098
1099         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE)
1100                 file_offset = record->leaf.base.key;
1101         else
1102                 file_offset = record->leaf.base.key - record->leaf.data_len;
1103         r = (file_offset < ip->save_trunc_off);
1104         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1105                 if (ip->save_trunc_off <= record->leaf.base.key)
1106                         ip->save_trunc_off = record->leaf.base.key + 1;
1107         } else {
1108                 if (ip->save_trunc_off < record->leaf.base.key)
1109                         ip->save_trunc_off = record->leaf.base.key;
1110         }
1111         return(r);
1112 }
1113
1114 /*
1115  * Backend code.  Sync a record to the media.
1116  */
1117 int
1118 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record)
1119 {
1120         hammer_transaction_t trans = cursor->trans;
1121         int64_t file_offset;
1122         int bytes;
1123         void *bdata;
1124         int error;
1125         int doprop;
1126
1127         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1128         KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE);
1129         KKASSERT(record->leaf.base.localization != 0);
1130
1131         /*
1132          * Any direct-write related to the record must complete before we
1133          * can sync the record to the on-disk media.
1134          */
1135         if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL))
1136                 hammer_io_direct_wait(record);
1137
1138         /*
1139          * If this is a bulk-data record placemarker there may be an existing
1140          * record on-disk, indicating a data overwrite.  If there is the
1141          * on-disk record must be deleted before we can insert our new record.
1142          *
1143          * We've synthesized this record and do not know what the create_tid
1144          * on-disk is, nor how much data it represents.
1145          *
1146          * Keep in mind that (key) for data records is (base_offset + len),
1147          * not (base_offset).  Also, we only want to get rid of on-disk
1148          * records since we are trying to sync our in-memory record, call
1149          * hammer_ip_delete_range() with truncating set to 1 to make sure
1150          * it skips in-memory records.
1151          *
1152          * It is ok for the lookup to return ENOENT.
1153          *
1154          * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have
1155          * to call hammer_ip_delete_range() or not.  This also means we must
1156          * update sync_trunc_off() as we write.
1157          */
1158         if (record->type == HAMMER_MEM_RECORD_DATA &&
1159             hammer_record_needs_overwrite_delete(record)) {
1160                 file_offset = record->leaf.base.key - record->leaf.data_len;
1161                 bytes = (record->leaf.data_len + HAMMER_BUFMASK) &
1162                         ~HAMMER_BUFMASK;
1163                 KKASSERT((file_offset & HAMMER_BUFMASK) == 0);
1164                 error = hammer_ip_delete_range(
1165                                 cursor, record->ip,
1166                                 file_offset, file_offset + bytes - 1,
1167                                 1);
1168                 if (error && error != ENOENT)
1169                         goto done;
1170         }
1171
1172         /*
1173          * If this is a general record there may be an on-disk version
1174          * that must be deleted before we can insert the new record.
1175          */
1176         if (record->type == HAMMER_MEM_RECORD_GENERAL) {
1177                 error = hammer_delete_general(cursor, record->ip,
1178                                               &record->leaf);
1179                 if (error && error != ENOENT)
1180                         goto done;
1181         }
1182
1183         /*
1184          * Setup the cursor.
1185          */
1186         hammer_normalize_cursor(cursor);
1187         cursor->key_beg = record->leaf.base;
1188         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1189         cursor->flags |= HAMMER_CURSOR_BACKEND;
1190         cursor->flags &= ~HAMMER_CURSOR_INSERT;
1191
1192         /*
1193          * Records can wind up on-media before the inode itself is on-media.
1194          * Flag the case.
1195          */
1196         record->ip->flags |= HAMMER_INODE_DONDISK;
1197
1198         /*
1199          * If we are deleting a directory entry an exact match must be
1200          * found on-disk.
1201          */
1202         if (record->type == HAMMER_MEM_RECORD_DEL) {
1203                 error = hammer_btree_lookup(cursor);
1204                 if (error == 0) {
1205                         KKASSERT(cursor->iprec == NULL);
1206                         error = hammer_ip_delete_record(cursor, record->ip,
1207                                                         trans->tid);
1208                         if (error == 0) {
1209                                 record->flags |= HAMMER_RECF_DELETED_BE |
1210                                                  HAMMER_RECF_COMMITTED;
1211                                 ++record->ip->rec_generation;
1212                         }
1213                 }
1214                 goto done;
1215         }
1216
1217         /*
1218          * We are inserting.
1219          *
1220          * Issue a lookup to position the cursor and locate the insertion
1221          * point.  The target key should not exist.  If we are creating a
1222          * directory entry we may have to iterate the low 32 bits of the
1223          * key to find an unused key.
1224          */
1225         hammer_sync_lock_sh(trans);
1226         cursor->flags |= HAMMER_CURSOR_INSERT;
1227         error = hammer_btree_lookup(cursor);
1228         if (hammer_debug_inode)
1229                 kprintf("DOINSERT LOOKUP %d\n", error);
1230         if (error == 0) {
1231                 kprintf("hammer_ip_sync_record: duplicate rec "
1232                         "at (%016llx)\n", (long long)record->leaf.base.key);
1233                 if (hammer_debug_critical)
1234                         Debugger("duplicate record1");
1235                 error = EIO;
1236         }
1237 #if 0
1238         if (record->type == HAMMER_MEM_RECORD_DATA)
1239                 kprintf("sync_record  %016llx ---------------- %016llx %d\n",
1240                         record->leaf.base.key - record->leaf.data_len,
1241                         record->leaf.data_offset, error);
1242 #endif
1243
1244         if (error != ENOENT)
1245                 goto done_unlock;
1246
1247         /*
1248          * Allocate the record and data.  The result buffers will be
1249          * marked as being modified and further calls to
1250          * hammer_modify_buffer() will result in unneeded UNDO records.
1251          *
1252          * Support zero-fill records (data == NULL and data_len != 0)
1253          */
1254         if (record->type == HAMMER_MEM_RECORD_DATA) {
1255                 /*
1256                  * The data portion of a bulk-data record has already been
1257                  * committed to disk, we need only adjust the layer2
1258                  * statistics in the same transaction as our B-Tree insert.
1259                  */
1260                 KKASSERT(record->leaf.data_offset != 0);
1261                 error = hammer_blockmap_finalize(trans,
1262                                                  record->resv,
1263                                                  record->leaf.data_offset,
1264                                                  record->leaf.data_len);
1265         } else if (record->data && record->leaf.data_len) {
1266                 /*
1267                  * Wholely cached record, with data.  Allocate the data.
1268                  */
1269                 bdata = hammer_alloc_data(trans, record->leaf.data_len,
1270                                           record->leaf.base.rec_type,
1271                                           &record->leaf.data_offset,
1272                                           &cursor->data_buffer,
1273                                           0, &error);
1274                 if (bdata == NULL)
1275                         goto done_unlock;
1276                 hammer_crc_set_leaf(record->data, &record->leaf);
1277                 hammer_modify_buffer(trans, cursor->data_buffer, NULL, 0);
1278                 bcopy(record->data, bdata, record->leaf.data_len);
1279                 hammer_modify_buffer_done(cursor->data_buffer);
1280         } else {
1281                 /*
1282                  * Wholely cached record, without data.
1283                  */
1284                 record->leaf.data_offset = 0;
1285                 record->leaf.data_crc = 0;
1286         }
1287
1288         error = hammer_btree_insert(cursor, &record->leaf, &doprop);
1289         if (hammer_debug_inode && error) {
1290                 kprintf("BTREE INSERT error %d @ %016llx:%d key %016llx\n",
1291                         error,
1292                         (long long)cursor->node->node_offset,
1293                         cursor->index,
1294                         (long long)record->leaf.base.key);
1295         }
1296
1297         /*
1298          * Our record is on-disk and we normally mark the in-memory version
1299          * as having been committed (and not BE-deleted).
1300          *
1301          * If the record represented a directory deletion but we had to
1302          * sync a valid directory entry to disk due to dependancies,
1303          * we must convert the record to a covering delete so the
1304          * frontend does not have visibility on the synced entry.
1305          *
1306          * WARNING: cursor's leaf pointer may have changed after do_propagation
1307          *          returns!
1308          */
1309         if (error == 0) {
1310                 if (doprop) {
1311                         hammer_btree_do_propagation(cursor,
1312                                                     record->ip->pfsm,
1313                                                     &record->leaf);
1314                 }
1315                 if (record->flags & HAMMER_RECF_CONVERT_DELETE) {
1316                         /*
1317                          * Must convert deleted directory entry add
1318                          * to a directory entry delete.
1319                          */
1320                         KKASSERT(record->type == HAMMER_MEM_RECORD_ADD);
1321                         record->flags &= ~HAMMER_RECF_DELETED_FE;
1322                         record->type = HAMMER_MEM_RECORD_DEL;
1323                         KKASSERT(record->ip->obj_id == record->leaf.base.obj_id);
1324                         KKASSERT(record->flush_state == HAMMER_FST_FLUSH);
1325                         record->flags &= ~HAMMER_RECF_CONVERT_DELETE;
1326                         KKASSERT((record->flags & (HAMMER_RECF_COMMITTED |
1327                                                  HAMMER_RECF_DELETED_BE)) == 0);
1328                         /* converted record is not yet committed */
1329                         /* hammer_flush_record_done takes care of the rest */
1330                 } else {
1331                         /*
1332                          * Everything went fine and we are now done with
1333                          * this record.
1334                          */
1335                         record->flags |= HAMMER_RECF_COMMITTED;
1336                         ++record->ip->rec_generation;
1337                 }
1338         } else {
1339                 if (record->leaf.data_offset) {
1340                         hammer_blockmap_free(trans, record->leaf.data_offset,
1341                                              record->leaf.data_len);
1342                 }
1343         }
1344 done_unlock:
1345         hammer_sync_unlock(trans);
1346 done:
1347         return(error);
1348 }
1349
1350 /*
1351  * Add the record to the inode's rec_tree.  The low 32 bits of a directory
1352  * entry's key is used to deal with hash collisions in the upper 32 bits.
1353  * A unique 64 bit key is generated in-memory and may be regenerated a
1354  * second time when the directory record is flushed to the on-disk B-Tree.
1355  *
1356  * A referenced record is passed to this function.  This function
1357  * eats the reference.  If an error occurs the record will be deleted.
1358  *
1359  * A copy of the temporary record->data pointer provided by the caller
1360  * will be made.
1361  */
1362 int
1363 hammer_mem_add(hammer_record_t record)
1364 {
1365         hammer_mount_t hmp = record->ip->hmp;
1366
1367         /*
1368          * Make a private copy of record->data
1369          */
1370         if (record->data)
1371                 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA);
1372
1373         /*
1374          * Insert into the RB tree.  A unique key should have already
1375          * been selected if this is a directory entry.
1376          */
1377         if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) {
1378                 record->flags |= HAMMER_RECF_DELETED_FE;
1379                 hammer_rel_mem_record(record);
1380                 return (EEXIST);
1381         }
1382         ++hmp->count_newrecords;
1383         ++hmp->rsv_recs;
1384         ++record->ip->rsv_recs;
1385         record->ip->hmp->rsv_databytes += record->leaf.data_len;
1386         record->flags |= HAMMER_RECF_ONRBTREE;
1387         hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY);
1388         hammer_rel_mem_record(record);
1389         return(0);
1390 }
1391
1392 /************************************************************************
1393  *                   HAMMER INODE MERGED-RECORD FUNCTIONS               *
1394  ************************************************************************
1395  *
1396  * These functions augment the B-Tree scanning functions in hammer_btree.c
1397  * by merging in-memory records with on-disk records.
1398  */
1399
1400 /*
1401  * Locate a particular record either in-memory or on-disk.
1402  *
1403  * NOTE: This is basically a standalone routine, hammer_ip_next() may
1404  * NOT be called to iterate results.
1405  */
1406 int
1407 hammer_ip_lookup(hammer_cursor_t cursor)
1408 {
1409         int error;
1410
1411         /*
1412          * If the element is in-memory return it without searching the
1413          * on-disk B-Tree
1414          */
1415         KKASSERT(cursor->ip);
1416         error = hammer_mem_lookup(cursor);
1417         if (error == 0) {
1418                 cursor->leaf = &cursor->iprec->leaf;
1419                 return(error);
1420         }
1421         if (error != ENOENT)
1422                 return(error);
1423
1424         /*
1425          * If the inode has on-disk components search the on-disk B-Tree.
1426          */
1427         if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0)
1428                 return(error);
1429         error = hammer_btree_lookup(cursor);
1430         if (error == 0)
1431                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1432         return(error);
1433 }
1434
1435 /*
1436  * Helper for hammer_ip_first()/hammer_ip_next()
1437  *
1438  * NOTE: Both ATEDISK and DISKEOF will be set the same.  This sets up
1439  * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek
1440  * state if hammer_ip_next() needs to re-seek.
1441  */
1442 static __inline
1443 int
1444 _hammer_ip_seek_btree(hammer_cursor_t cursor)
1445 {
1446         hammer_inode_t ip = cursor->ip;
1447         int error;
1448
1449         if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) {
1450                 error = hammer_btree_lookup(cursor);
1451                 if (error == ENOENT || error == EDEADLK) {
1452                         if (hammer_debug_general & 0x2000) {
1453                                 kprintf("error %d node %p %016llx index %d\n",
1454                                         error, cursor->node,
1455                                         (long long)cursor->node->node_offset,
1456                                         cursor->index);
1457                         }
1458                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1459                         error = hammer_btree_iterate(cursor);
1460                 }
1461                 if (error == 0) {
1462                         cursor->flags &= ~(HAMMER_CURSOR_DISKEOF |
1463                                            HAMMER_CURSOR_ATEDISK);
1464                 } else {
1465                         cursor->flags |= HAMMER_CURSOR_DISKEOF |
1466                                          HAMMER_CURSOR_ATEDISK;
1467                         if (error == ENOENT)
1468                                 error = 0;
1469                 }
1470         } else {
1471                 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK;
1472                 error = 0;
1473         }
1474         return(error);
1475 }
1476
1477 /*
1478  * Helper for hammer_ip_next()
1479  *
1480  * The caller has determined that the media cursor is further along than the
1481  * memory cursor and must be reseeked after a generation number change.
1482  */
1483 static
1484 int
1485 _hammer_ip_reseek(hammer_cursor_t cursor)
1486 {
1487         struct hammer_base_elm save;
1488         hammer_btree_elm_t elm;
1489         int error;
1490         int r;
1491         int again = 0;
1492
1493         /*
1494          * Do the re-seek.
1495          */
1496         kprintf("HAMMER: Debug: re-seeked during scan @ino=%016llx\n",
1497                 (long long)cursor->ip->obj_id);
1498         save = cursor->key_beg;
1499         cursor->key_beg = cursor->iprec->leaf.base;
1500         error = _hammer_ip_seek_btree(cursor);
1501         KKASSERT(error == 0);
1502         cursor->key_beg = save;
1503
1504         /*
1505          * If the memory record was previous returned to
1506          * the caller and the media record matches
1507          * (-1/+1: only create_tid differs), then iterate
1508          * the media record to avoid a double result.
1509          */
1510         if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 &&
1511             (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) {
1512                 elm = &cursor->node->ondisk->elms[cursor->index];
1513                 r = hammer_btree_cmp(&elm->base,
1514                                      &cursor->iprec->leaf.base);
1515                 if (cursor->flags & HAMMER_CURSOR_ASOF) {
1516                         if (r >= -1 && r <= 1) {
1517                                 kprintf("HAMMER: Debug: iterated after "
1518                                         "re-seek (asof r=%d)\n", r);
1519                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1520                                 again = 1;
1521                         }
1522                 } else {
1523                         if (r == 0) {
1524                                 kprintf("HAMMER: Debug: iterated after "
1525                                         "re-seek\n");
1526                                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1527                                 again = 1;
1528                         }
1529                 }
1530         }
1531         return(again);
1532 }
1533
1534 /*
1535  * Locate the first record within the cursor's key_beg/key_end range,
1536  * restricted to a particular inode.  0 is returned on success, ENOENT
1537  * if no records matched the requested range, or some other error.
1538  *
1539  * When 0 is returned hammer_ip_next() may be used to iterate additional
1540  * records within the requested range.
1541  *
1542  * This function can return EDEADLK, requiring the caller to terminate
1543  * the cursor and try again.
1544  */
1545
1546 int
1547 hammer_ip_first(hammer_cursor_t cursor)
1548 {
1549         hammer_inode_t ip __debugvar = cursor->ip;
1550         int error;
1551
1552         KKASSERT(ip != NULL);
1553
1554         /*
1555          * Clean up fields and setup for merged scan
1556          */
1557         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1558
1559         /*
1560          * Search the in-memory record list (Red-Black tree).  Unlike the
1561          * B-Tree search, mem_first checks for records in the range.
1562          *
1563          * This function will setup both ATEMEM and MEMEOF properly for
1564          * the ip iteration.  ATEMEM will be set if MEMEOF is set.
1565          */
1566         hammer_mem_first(cursor);
1567
1568         /*
1569          * Detect generation changes during blockages, including
1570          * blockages which occur on the initial btree search.
1571          */
1572         cursor->rec_generation = cursor->ip->rec_generation;
1573
1574         /*
1575          * Initial search and result
1576          */
1577         error = _hammer_ip_seek_btree(cursor);
1578         if (error == 0)
1579                 error = hammer_ip_next(cursor);
1580
1581         return (error);
1582 }
1583
1584 /*
1585  * Retrieve the next record in a merged iteration within the bounds of the
1586  * cursor.  This call may be made multiple times after the cursor has been
1587  * initially searched with hammer_ip_first().
1588  *
1589  * There are numerous special cases in this code to deal with races between
1590  * in-memory records and on-media records.
1591  *
1592  * 0 is returned on success, ENOENT if no further records match the
1593  * requested range, or some other error code is returned.
1594  */
1595 int
1596 hammer_ip_next(hammer_cursor_t cursor)
1597 {
1598         hammer_btree_elm_t elm;
1599         hammer_record_t rec;
1600         hammer_record_t tmprec;
1601         int error;
1602         int r;
1603
1604 again:
1605         /*
1606          * Get the next on-disk record
1607          *
1608          * NOTE: If we deleted the last on-disk record we had scanned
1609          *       ATEDISK will be clear and RETEST will be set, forcing
1610          *       a call to iterate.  The fact that ATEDISK is clear causes
1611          *       iterate to re-test the 'current' element.  If ATEDISK is
1612          *       set, iterate will skip the 'current' element.
1613          */
1614         error = 0;
1615         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
1616                 if (cursor->flags & (HAMMER_CURSOR_ATEDISK |
1617                                      HAMMER_CURSOR_RETEST)) {
1618                         error = hammer_btree_iterate(cursor);
1619                         cursor->flags &= ~HAMMER_CURSOR_RETEST;
1620                         if (error == 0) {
1621                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
1622                                 hammer_cache_node(&cursor->ip->cache[1],
1623                                                   cursor->node);
1624                         } else if (error == ENOENT) {
1625                                 cursor->flags |= HAMMER_CURSOR_DISKEOF |
1626                                                  HAMMER_CURSOR_ATEDISK;
1627                                 error = 0;
1628                         }
1629                 }
1630         }
1631
1632         /*
1633          * If the generation changed the backend has deleted or committed
1634          * one or more memory records since our last check.
1635          *
1636          * When this case occurs if the disk cursor is > current memory record
1637          * or the disk cursor is at EOF, we must re-seek the disk-cursor.
1638          * Since the cursor is ahead it must have not yet been eaten (if
1639          * not at eof anyway). (XXX data offset case?)
1640          *
1641          * NOTE: we are not doing a full check here.  That will be handled
1642          * later on.
1643          *
1644          * If we have exhausted all memory records we do not have to do any
1645          * further seeks.
1646          */
1647         while (cursor->rec_generation != cursor->ip->rec_generation &&
1648                error == 0
1649         ) {
1650                 kprintf("HAMMER: Debug: generation changed during scan @ino=%016llx\n", (long long)cursor->ip->obj_id);
1651                 cursor->rec_generation = cursor->ip->rec_generation;
1652                 if (cursor->flags & HAMMER_CURSOR_MEMEOF)
1653                         break;
1654                 if (cursor->flags & HAMMER_CURSOR_DISKEOF) {
1655                         r = 1;
1656                 } else {
1657                         KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0);
1658                         elm = &cursor->node->ondisk->elms[cursor->index];
1659                         r = hammer_btree_cmp(&elm->base,
1660                                              &cursor->iprec->leaf.base);
1661                 }
1662
1663                 /*
1664                  * Do we re-seek the media cursor?
1665                  */
1666                 if (r > 0) {
1667                         if (_hammer_ip_reseek(cursor))
1668                                 goto again;
1669                 }
1670         }
1671
1672         /*
1673          * We can now safely get the next in-memory record.  We cannot
1674          * block here.
1675          *
1676          * hammer_rec_scan_cmp:  Is the record still in our general range,
1677          *                       (non-inclusive of snapshot exclusions)?
1678          * hammer_rec_scan_callback: Is the record in our snapshot?
1679          */
1680         tmprec = NULL;
1681         if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1682                 /*
1683                  * If the current memory record was eaten then get the next
1684                  * one.  Stale records are skipped.
1685                  */
1686                 if (cursor->flags & HAMMER_CURSOR_ATEMEM) {
1687                         tmprec = cursor->iprec;
1688                         cursor->iprec = NULL;
1689                         rec = hammer_rec_rb_tree_RB_NEXT(tmprec);
1690                         while (rec) {
1691                                 if (hammer_rec_scan_cmp(rec, cursor) != 0)
1692                                         break;
1693                                 if (hammer_rec_scan_callback(rec, cursor) != 0)
1694                                         break;
1695                                 rec = hammer_rec_rb_tree_RB_NEXT(rec);
1696                         }
1697                         if (cursor->iprec) {
1698                                 KKASSERT(cursor->iprec == rec);
1699                                 cursor->flags &= ~HAMMER_CURSOR_ATEMEM;
1700                         } else {
1701                                 cursor->flags |= HAMMER_CURSOR_MEMEOF;
1702                         }
1703                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1704                 }
1705         }
1706
1707         /*
1708          * MEMORY RECORD VALIDITY TEST
1709          *
1710          * (We still can't block, which is why tmprec is being held so
1711          * long).
1712          *
1713          * If the memory record is no longer valid we skip it.  It may
1714          * have been deleted by the frontend.  If it was deleted or
1715          * committed by the backend the generation change re-seeked the
1716          * disk cursor and the record will be present there.
1717          */
1718         if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) {
1719                 KKASSERT(cursor->iprec);
1720                 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0);
1721                 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) {
1722                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1723                         if (tmprec)
1724                                 hammer_rel_mem_record(tmprec);
1725                         goto again;
1726                 }
1727         }
1728         if (tmprec)
1729                 hammer_rel_mem_record(tmprec);
1730
1731         /*
1732          * Extract either the disk or memory record depending on their
1733          * relative position.
1734          */
1735         error = 0;
1736         switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) {
1737         case 0:
1738                 /*
1739                  * Both entries valid.   Compare the entries and nominally
1740                  * return the first one in the sort order.  Numerous cases
1741                  * require special attention, however.
1742                  */
1743                 elm = &cursor->node->ondisk->elms[cursor->index];
1744                 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base);
1745
1746                 /*
1747                  * If the two entries differ only by their key (-2/2) or
1748                  * create_tid (-1/1), and are DATA records, we may have a
1749                  * nominal match.  We have to calculate the base file
1750                  * offset of the data.
1751                  */
1752                 if (r <= 2 && r >= -2 && r != 0 &&
1753                     cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE &&
1754                     cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1755                         int64_t base1 = elm->leaf.base.key - elm->leaf.data_len;
1756                         int64_t base2 = cursor->iprec->leaf.base.key -
1757                                         cursor->iprec->leaf.data_len;
1758                         if (base1 == base2)
1759                                 r = 0;
1760                 }
1761
1762                 if (r < 0) {
1763                         error = hammer_btree_extract(cursor,
1764                                                      HAMMER_CURSOR_GET_LEAF);
1765                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1766                         cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1767                         break;
1768                 }
1769
1770                 /*
1771                  * If the entries match exactly the memory entry is either
1772                  * an on-disk directory entry deletion or a bulk data
1773                  * overwrite.  If it is a directory entry deletion we eat
1774                  * both entries.
1775                  *
1776                  * For the bulk-data overwrite case it is possible to have
1777                  * visibility into both, which simply means the syncer
1778                  * hasn't gotten around to doing the delete+insert sequence
1779                  * on the B-Tree.  Use the memory entry and throw away the
1780                  * on-disk entry.
1781                  *
1782                  * If the in-memory record is not either of these we
1783                  * probably caught the syncer while it was syncing it to
1784                  * the media.  Since we hold a shared lock on the cursor,
1785                  * the in-memory record had better be marked deleted at
1786                  * this point.
1787                  */
1788                 if (r == 0) {
1789                         if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) {
1790                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1791                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1792                                         cursor->flags |= HAMMER_CURSOR_ATEMEM;
1793                                         goto again;
1794                                 }
1795                         } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) {
1796                                 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1797                                         cursor->flags |= HAMMER_CURSOR_ATEDISK;
1798                                 }
1799                                 /* fall through to memory entry */
1800                         } else {
1801                                 panic("hammer_ip_next: duplicate mem/b-tree entry %p %d %08x", cursor->iprec, cursor->iprec->type, cursor->iprec->flags);
1802                                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1803                                 goto again;
1804                         }
1805                 }
1806                 /* fall through to the memory entry */
1807         case HAMMER_CURSOR_ATEDISK:
1808                 /*
1809                  * Only the memory entry is valid.
1810                  */
1811                 cursor->leaf = &cursor->iprec->leaf;
1812                 cursor->flags |= HAMMER_CURSOR_ATEMEM;
1813                 cursor->flags |= HAMMER_CURSOR_LASTWASMEM;
1814
1815                 /*
1816                  * If the memory entry is an on-disk deletion we should have
1817                  * also had found a B-Tree record.  If the backend beat us
1818                  * to it it would have interlocked the cursor and we should
1819                  * have seen the in-memory record marked DELETED_FE.
1820                  */
1821                 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL &&
1822                     (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) {
1823                         panic("hammer_ip_next: del-on-disk with no b-tree entry iprec %p flags %08x", cursor->iprec, cursor->iprec->flags);
1824                 }
1825                 break;
1826         case HAMMER_CURSOR_ATEMEM:
1827                 /*
1828                  * Only the disk entry is valid
1829                  */
1830                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
1831                 cursor->flags |= HAMMER_CURSOR_ATEDISK;
1832                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1833                 break;
1834         default:
1835                 /*
1836                  * Neither entry is valid
1837                  *
1838                  * XXX error not set properly
1839                  */
1840                 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM;
1841                 cursor->leaf = NULL;
1842                 error = ENOENT;
1843                 break;
1844         }
1845         return(error);
1846 }
1847
1848 /*
1849  * Resolve the cursor->data pointer for the current cursor position in
1850  * a merged iteration.
1851  */
1852 int
1853 hammer_ip_resolve_data(hammer_cursor_t cursor)
1854 {
1855         hammer_record_t record;
1856         int error;
1857
1858         if (hammer_cursor_inmem(cursor)) {
1859                 /*
1860                  * The data associated with an in-memory record is usually
1861                  * kmalloced, but reserve-ahead data records will have an
1862                  * on-disk reference.
1863                  *
1864                  * NOTE: Reserve-ahead data records must be handled in the
1865                  * context of the related high level buffer cache buffer
1866                  * to interlock against async writes.
1867                  */
1868                 record = cursor->iprec;
1869                 cursor->data = record->data;
1870                 error = 0;
1871                 if (cursor->data == NULL) {
1872                         KKASSERT(record->leaf.base.rec_type ==
1873                                  HAMMER_RECTYPE_DATA);
1874                         cursor->data = hammer_bread_ext(cursor->trans->hmp,
1875                                                     record->leaf.data_offset,
1876                                                     record->leaf.data_len,
1877                                                     &error,
1878                                                     &cursor->data_buffer);
1879                 }
1880         } else {
1881                 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf;
1882                 error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_DATA);
1883         }
1884         return(error);
1885 }
1886
1887 /*
1888  * Backend truncation / record replacement - delete records in range.
1889  *
1890  * Delete all records within the specified range for inode ip.  In-memory
1891  * records still associated with the frontend are ignored.
1892  *
1893  * If truncating is non-zero in-memory records associated with the back-end
1894  * are ignored.  If truncating is > 1 we can return EWOULDBLOCK.
1895  *
1896  * NOTES:
1897  *
1898  *      * An unaligned range will cause new records to be added to cover
1899  *        the edge cases. (XXX not implemented yet).
1900  *
1901  *      * Replacement via reservations (see hammer_ip_sync_record_cursor())
1902  *        also do not deal with unaligned ranges.
1903  *
1904  *      * ran_end is inclusive (e.g. 0,1023 instead of 0,1024).
1905  *
1906  *      * Record keys for regular file data have to be special-cased since
1907  *        they indicate the end of the range (key = base + bytes).
1908  *
1909  *      * This function may be asked to delete ridiculously huge ranges, for
1910  *        example if someone truncates or removes a 1TB regular file.  We
1911  *        must be very careful on restarts and we may have to stop w/
1912  *        EWOULDBLOCK to avoid blowing out the buffer cache.
1913  */
1914 int
1915 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip,
1916                        int64_t ran_beg, int64_t ran_end, int truncating)
1917 {
1918         hammer_transaction_t trans = cursor->trans;
1919         hammer_btree_leaf_elm_t leaf;
1920         int error;
1921         int64_t off;
1922         int64_t tmp64;
1923
1924 #if 0
1925         kprintf("delete_range %p %016llx-%016llx\n", ip, ran_beg, ran_end);
1926 #endif
1927
1928         KKASSERT(trans->type == HAMMER_TRANS_FLS);
1929 retry:
1930         hammer_normalize_cursor(cursor);
1931         cursor->key_beg.localization = ip->obj_localization +
1932                                        HAMMER_LOCALIZE_MISC;
1933         cursor->key_beg.obj_id = ip->obj_id;
1934         cursor->key_beg.create_tid = 0;
1935         cursor->key_beg.delete_tid = 0;
1936         cursor->key_beg.obj_type = 0;
1937
1938         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1939                 cursor->key_beg.key = ran_beg;
1940                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB;
1941         } else {
1942                 /*
1943                  * The key in the B-Tree is (base+bytes), so the first possible
1944                  * matching key is ran_beg + 1.
1945                  */
1946                 cursor->key_beg.key = ran_beg + 1;
1947                 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA;
1948         }
1949
1950         cursor->key_end = cursor->key_beg;
1951         if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) {
1952                 cursor->key_end.key = ran_end;
1953         } else {
1954                 tmp64 = ran_end + MAXPHYS + 1;  /* work around GCC-4 bug */
1955                 if (tmp64 < ran_end)
1956                         cursor->key_end.key = 0x7FFFFFFFFFFFFFFFLL;
1957                 else
1958                         cursor->key_end.key = ran_end + MAXPHYS + 1;
1959         }
1960
1961         cursor->asof = ip->obj_asof;
1962         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
1963         cursor->flags |= HAMMER_CURSOR_ASOF;
1964         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
1965         cursor->flags |= HAMMER_CURSOR_BACKEND;
1966         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE;
1967
1968         error = hammer_ip_first(cursor);
1969
1970         /*
1971          * Iterate through matching records and mark them as deleted.
1972          */
1973         while (error == 0) {
1974                 leaf = cursor->leaf;
1975
1976                 KKASSERT(leaf->base.delete_tid == 0);
1977                 KKASSERT(leaf->base.obj_id == ip->obj_id);
1978
1979                 /*
1980                  * There may be overlap cases for regular file data.  Also
1981                  * remember the key for a regular file record is (base + len),
1982                  * NOT (base).
1983                  *
1984                  * Note that due to duplicates (mem & media) allowed by
1985                  * DELETE_VISIBILITY, off can wind up less then ran_beg.
1986                  */
1987                 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) {
1988                         off = leaf->base.key - leaf->data_len;
1989                         /*
1990                          * Check the left edge case.  We currently do not
1991                          * split existing records.
1992                          */
1993                         if (off < ran_beg && leaf->base.key > ran_beg) {
1994                                 panic("hammer left edge case %016llx %d\n",
1995                                         (long long)leaf->base.key,
1996                                         leaf->data_len);
1997                         }
1998
1999                         /*
2000                          * Check the right edge case.  Note that the
2001                          * record can be completely out of bounds, which
2002                          * terminates the search.
2003                          *
2004                          * base->key is exclusive of the right edge while
2005                          * ran_end is inclusive of the right edge.  The
2006                          * (key - data_len) left boundary is inclusive.
2007                          *
2008                          * XXX theory-check this test at some point, are
2009                          * we missing a + 1 somewhere?  Note that ran_end
2010                          * could overflow.
2011                          */
2012                         if (leaf->base.key - 1 > ran_end) {
2013                                 if (leaf->base.key - leaf->data_len > ran_end)
2014                                         break;
2015                                 panic("hammer right edge case\n");
2016                         }
2017                 } else {
2018                         off = leaf->base.key;
2019                 }
2020
2021                 /*
2022                  * Delete the record.  When truncating we do not delete
2023                  * in-memory (data) records because they represent data
2024                  * written after the truncation.
2025                  *
2026                  * This will also physically destroy the B-Tree entry and
2027                  * data if the retention policy dictates.  The function
2028                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2029                  * to retest the new 'current' element.
2030                  */
2031                 if (truncating == 0 || hammer_cursor_ondisk(cursor)) {
2032                         error = hammer_ip_delete_record(cursor, ip, trans->tid);
2033                         /*
2034                          * If we have built up too many meta-buffers we risk
2035                          * deadlocking the kernel and must stop.  This can
2036                          * occur when deleting ridiculously huge files.
2037                          * sync_trunc_off is updated so the next cycle does
2038                          * not re-iterate records we have already deleted.
2039                          *
2040                          * This is only done with formal truncations.
2041                          */
2042                         if (truncating > 1 && error == 0 &&
2043                             hammer_flusher_meta_limit(ip->hmp)) {
2044                                 ip->sync_trunc_off = off;
2045                                 error = EWOULDBLOCK;
2046                         }
2047                 }
2048                 if (error)
2049                         break;
2050                 ran_beg = off;  /* for restart */
2051                 error = hammer_ip_next(cursor);
2052         }
2053         if (cursor->node)
2054                 hammer_cache_node(&ip->cache[1], cursor->node);
2055
2056         if (error == EDEADLK) {
2057                 hammer_done_cursor(cursor);
2058                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2059                 if (error == 0)
2060                         goto retry;
2061         }
2062         if (error == ENOENT)
2063                 error = 0;
2064         return(error);
2065 }
2066
2067 /*
2068  * This backend function deletes the specified record on-disk, similar to
2069  * delete_range but for a specific record.  Unlike the exact deletions
2070  * used when deleting a directory entry this function uses an ASOF search
2071  * like delete_range.
2072  *
2073  * This function may be called with ip->obj_asof set for a slave snapshot,
2074  * so don't use it.  We always delete non-historical records only.
2075  */
2076 static int
2077 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip,
2078                       hammer_btree_leaf_elm_t leaf)
2079 {
2080         hammer_transaction_t trans = cursor->trans;
2081         int error;
2082
2083         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2084 retry:
2085         hammer_normalize_cursor(cursor);
2086         cursor->key_beg = leaf->base;
2087         cursor->asof = HAMMER_MAX_TID;
2088         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2089         cursor->flags |= HAMMER_CURSOR_ASOF;
2090         cursor->flags |= HAMMER_CURSOR_BACKEND;
2091         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2092
2093         error = hammer_btree_lookup(cursor);
2094         if (error == 0) {
2095                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2096         }
2097         if (error == EDEADLK) {
2098                 hammer_done_cursor(cursor);
2099                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2100                 if (error == 0)
2101                         goto retry;
2102         }
2103         return(error);
2104 }
2105
2106 /*
2107  * This function deletes remaining auxillary records when an inode is
2108  * being deleted.  This function explicitly does not delete the
2109  * inode record, directory entry, data, or db records.  Those must be
2110  * properly disposed of prior to this call.
2111  */
2112 int
2113 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp)
2114 {
2115         hammer_transaction_t trans = cursor->trans;
2116         hammer_btree_leaf_elm_t leaf;
2117         int error;
2118
2119         KKASSERT(trans->type == HAMMER_TRANS_FLS);
2120 retry:
2121         hammer_normalize_cursor(cursor);
2122         cursor->key_beg.localization = ip->obj_localization +
2123                                        HAMMER_LOCALIZE_MISC;
2124         cursor->key_beg.obj_id = ip->obj_id;
2125         cursor->key_beg.create_tid = 0;
2126         cursor->key_beg.delete_tid = 0;
2127         cursor->key_beg.obj_type = 0;
2128         cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START;
2129         cursor->key_beg.key = HAMMER_MIN_KEY;
2130
2131         cursor->key_end = cursor->key_beg;
2132         cursor->key_end.rec_type = HAMMER_RECTYPE_MAX;
2133         cursor->key_end.key = HAMMER_MAX_KEY;
2134
2135         cursor->asof = ip->obj_asof;
2136         cursor->flags &= ~HAMMER_CURSOR_INITMASK;
2137         cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2138         cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY;
2139         cursor->flags |= HAMMER_CURSOR_BACKEND;
2140
2141         error = hammer_ip_first(cursor);
2142
2143         /*
2144          * Iterate through matching records and mark them as deleted.
2145          */
2146         while (error == 0) {
2147                 leaf = cursor->leaf;
2148
2149                 KKASSERT(leaf->base.delete_tid == 0);
2150
2151                 /*
2152                  * Mark the record and B-Tree entry as deleted.  This will
2153                  * also physically delete the B-Tree entry, record, and
2154                  * data if the retention policy dictates.  The function
2155                  * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next()
2156                  * to retest the new 'current' element.
2157                  *
2158                  * Directory entries (and delete-on-disk directory entries)
2159                  * must be synced and cannot be deleted.
2160                  */
2161                 error = hammer_ip_delete_record(cursor, ip, trans->tid);
2162                 ++*countp;
2163                 if (error)
2164                         break;
2165                 error = hammer_ip_next(cursor);
2166         }
2167         if (cursor->node)
2168                 hammer_cache_node(&ip->cache[1], cursor->node);
2169         if (error == EDEADLK) {
2170                 hammer_done_cursor(cursor);
2171                 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip);
2172                 if (error == 0)
2173                         goto retry;
2174         }
2175         if (error == ENOENT)
2176                 error = 0;
2177         return(error);
2178 }
2179
2180 /*
2181  * Delete the record at the current cursor.  On success the cursor will
2182  * be positioned appropriately for an iteration but may no longer be at
2183  * a leaf node.
2184  *
2185  * This routine is only called from the backend.
2186  *
2187  * NOTE: This can return EDEADLK, requiring the caller to terminate the
2188  * cursor and retry.
2189  */
2190 int
2191 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip,
2192                         hammer_tid_t tid)
2193 {
2194         hammer_record_t iprec;
2195         hammer_mount_t hmp;
2196         int error;
2197
2198         KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND);
2199         KKASSERT(tid != 0);
2200         hmp = cursor->node->hmp;
2201
2202         /*
2203          * In-memory (unsynchronized) records can simply be freed.  This
2204          * only occurs in range iterations since all other records are
2205          * individually synchronized.  Thus there should be no confusion with
2206          * the interlock.
2207          *
2208          * An in-memory record may be deleted before being committed to disk,
2209          * but could have been accessed in the mean time.  The reservation
2210          * code will deal with the case.
2211          */
2212         if (hammer_cursor_inmem(cursor)) {
2213                 iprec = cursor->iprec;
2214                 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0);
2215                 iprec->flags |= HAMMER_RECF_DELETED_FE;
2216                 iprec->flags |= HAMMER_RECF_DELETED_BE;
2217                 KKASSERT(iprec->ip == ip);
2218                 ++ip->rec_generation;
2219                 return(0);
2220         }
2221
2222         /*
2223          * On-disk records are marked as deleted by updating their delete_tid.
2224          * This does not effect their position in the B-Tree (which is based
2225          * on their create_tid).
2226          *
2227          * Frontend B-Tree operations track inodes so we tell
2228          * hammer_delete_at_cursor() not to.
2229          */
2230         error = hammer_btree_extract(cursor, HAMMER_CURSOR_GET_LEAF);
2231
2232         if (error == 0) {
2233                 error = hammer_delete_at_cursor(
2234                                 cursor,
2235                                 HAMMER_DELETE_ADJUST | hammer_nohistory(ip),
2236                                 cursor->trans->tid,
2237                                 cursor->trans->time32,
2238                                 0, NULL);
2239         }
2240         return(error);
2241 }
2242
2243 /*
2244  * Used to write a generic record w/optional data to the media b-tree
2245  * when no inode context is available.  Used by the mirroring and
2246  * snapshot code.
2247  *
2248  * Caller must set cursor->key_beg to leaf->base.  The cursor must be
2249  * flagged for backend operation and not flagged ASOF (since we are
2250  * doing an insertion).
2251  *
2252  * This function will acquire the appropriate sync lock and will set
2253  * the cursor insertion flag for the operation, do the btree lookup,
2254  * and the insertion, and clear the insertion flag and sync lock before
2255  * returning.  The cursor state will be such that the caller can continue
2256  * scanning (used by the mirroring code).
2257  *
2258  * mode: HAMMER_CREATE_MODE_UMIRROR     copyin data, check crc
2259  *       HAMMER_CREATE_MODE_SYS         bcopy data, generate crc
2260  *
2261  * NOTE: EDEADLK can be returned.  The caller must do deadlock handling and
2262  *                retry.
2263  *
2264  *       EALREADY can be returned if the record already exists (WARNING,
2265  *                because ASOF cannot be used no check is made for illegal
2266  *                duplicates).
2267  *
2268  * NOTE: Do not use the function for normal inode-related records as this
2269  *       functions goes directly to the media and is not integrated with
2270  *       in-memory records.
2271  */
2272 int
2273 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf,
2274                         void *udata, int mode)
2275 {
2276         hammer_transaction_t trans;
2277         hammer_buffer_t data_buffer;
2278         hammer_off_t ndata_offset;
2279         hammer_tid_t high_tid;
2280         void *ndata;
2281         int error;
2282         int doprop;
2283
2284         trans = cursor->trans;
2285         data_buffer = NULL;
2286         ndata_offset = 0;
2287         doprop = 0;
2288
2289         KKASSERT((cursor->flags &
2290                   (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) ==
2291                   (HAMMER_CURSOR_BACKEND));
2292
2293         hammer_sync_lock_sh(trans);
2294
2295         if (leaf->data_len) {
2296                 ndata = hammer_alloc_data(trans, leaf->data_len,
2297                                           leaf->base.rec_type,
2298                                           &ndata_offset, &data_buffer,
2299                                           0, &error);
2300                 if (ndata == NULL) {
2301                         hammer_sync_unlock(trans);
2302                         return (error);
2303                 }
2304                 leaf->data_offset = ndata_offset;
2305                 hammer_modify_buffer(trans, data_buffer, NULL, 0);
2306
2307                 switch(mode) {
2308                 case HAMMER_CREATE_MODE_UMIRROR:
2309                         error = copyin(udata, ndata, leaf->data_len);
2310                         if (error == 0) {
2311                                 if (hammer_crc_test_leaf(ndata, leaf) == 0) {
2312                                         kprintf("data crc mismatch on pipe\n");
2313                                         error = EINVAL;
2314                                 } else {
2315                                         error = hammer_cursor_localize_data(
2316                                                         ndata, leaf);
2317                                 }
2318                         }
2319                         break;
2320                 case HAMMER_CREATE_MODE_SYS:
2321                         bcopy(udata, ndata, leaf->data_len);
2322                         error = 0;
2323                         hammer_crc_set_leaf(ndata, leaf);
2324                         break;
2325                 default:
2326                         panic("hammer: hammer_create_at_cursor: bad mode %d",
2327                                 mode);
2328                         break; /* NOT REACHED */
2329                 }
2330                 hammer_modify_buffer_done(data_buffer);
2331         } else {
2332                 leaf->data_offset = 0;
2333                 error = 0;
2334                 ndata = NULL;
2335         }
2336         if (error)
2337                 goto failed;
2338
2339         /*
2340          * Do the insertion.  This can fail with a EDEADLK or EALREADY
2341          */
2342         cursor->flags |= HAMMER_CURSOR_INSERT;
2343         error = hammer_btree_lookup(cursor);
2344         if (error != ENOENT) {
2345                 if (error == 0)
2346                         error = EALREADY;
2347                 goto failed;
2348         }
2349         error = hammer_btree_insert(cursor, leaf, &doprop);
2350
2351         /*
2352          * Cursor is left on current element, we want to skip it now.
2353          * (in case the caller is scanning)
2354          */
2355         cursor->flags |= HAMMER_CURSOR_ATEDISK;
2356         cursor->flags &= ~HAMMER_CURSOR_INSERT;
2357
2358         /*
2359          * If the insertion happens to be creating (and not just replacing)
2360          * an inode we have to track it.
2361          */
2362         if (error == 0 &&
2363             leaf->base.rec_type == HAMMER_RECTYPE_INODE &&
2364             leaf->base.delete_tid == 0) {
2365                 hammer_modify_volume_field(trans, trans->rootvol,
2366                                            vol0_stat_inodes);
2367                 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes;
2368                 hammer_modify_volume_done(trans->rootvol);
2369         }
2370
2371         /*
2372          * vol0_next_tid must track the highest TID stored in the filesystem.
2373          * We do not need to generate undo for this update.
2374          */
2375         high_tid = leaf->base.create_tid;
2376         if (high_tid < leaf->base.delete_tid)
2377                 high_tid = leaf->base.delete_tid;
2378         if (trans->rootvol->ondisk->vol0_next_tid < high_tid) {
2379                 hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2380                 trans->rootvol->ondisk->vol0_next_tid = high_tid;
2381                 hammer_modify_volume_done(trans->rootvol);
2382         }
2383
2384         /*
2385          * WARNING!  cursor's leaf pointer may have changed after
2386          *           do_propagation returns.
2387          */
2388         if (error == 0 && doprop)
2389                 hammer_btree_do_propagation(cursor, NULL, leaf);
2390
2391 failed:
2392         /*
2393          * Cleanup
2394          */
2395         if (error && leaf->data_offset) {
2396                 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len);
2397
2398         }
2399         hammer_sync_unlock(trans);
2400         if (data_buffer)
2401                 hammer_rel_buffer(data_buffer, 0);
2402         return (error);
2403 }
2404
2405 /*
2406  * Delete the B-Tree element at the current cursor and do any necessary
2407  * mirror propagation.
2408  *
2409  * The cursor must be properly positioned for an iteration on return but
2410  * may be pointing at an internal element.
2411  *
2412  * An element can be un-deleted by passing a delete_tid of 0 with
2413  * HAMMER_DELETE_ADJUST.
2414  */
2415 int
2416 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags,
2417                         hammer_tid_t delete_tid, u_int32_t delete_ts,
2418                         int track, int64_t *stat_bytes)
2419 {
2420         struct hammer_btree_leaf_elm save_leaf;
2421         hammer_transaction_t trans;
2422         hammer_btree_leaf_elm_t leaf;
2423         hammer_node_t node;
2424         hammer_btree_elm_t elm;
2425         hammer_off_t data_offset;
2426         int32_t data_len;
2427         u_int16_t rec_type;
2428         int error;
2429         int icount;
2430         int doprop;
2431
2432         error = hammer_cursor_upgrade(cursor);
2433         if (error)
2434                 return(error);
2435
2436         trans = cursor->trans;
2437         node = cursor->node;
2438         elm = &node->ondisk->elms[cursor->index];
2439         leaf = &elm->leaf;
2440         KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD);
2441
2442         hammer_sync_lock_sh(trans);
2443         doprop = 0;
2444         icount = 0;
2445
2446         /*
2447          * Adjust the delete_tid.  Update the mirror_tid propagation field
2448          * as well.  delete_tid can be 0 (undelete -- used by mirroring).
2449          */
2450         if (delete_flags & HAMMER_DELETE_ADJUST) {
2451                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) {
2452                         if (elm->leaf.base.delete_tid == 0 && delete_tid)
2453                                 icount = -1;
2454                         if (elm->leaf.base.delete_tid && delete_tid == 0)
2455                                 icount = 1;
2456                 }
2457
2458                 hammer_modify_node(trans, node, elm, sizeof(*elm));
2459                 elm->leaf.base.delete_tid = delete_tid;
2460                 elm->leaf.delete_ts = delete_ts;
2461                 hammer_modify_node_done(node);
2462
2463                 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) {
2464                         hammer_modify_node_field(trans, node, mirror_tid);
2465                         node->ondisk->mirror_tid = elm->leaf.base.delete_tid;
2466                         hammer_modify_node_done(node);
2467                         doprop = 1;
2468                         if (hammer_debug_general & 0x0002) {
2469                                 kprintf("delete_at_cursor: propagate %016llx"
2470                                         " @%016llx\n",
2471                                         (long long)elm->leaf.base.delete_tid,
2472                                         (long long)node->node_offset);
2473                         }
2474                 }
2475
2476                 /*
2477                  * Adjust for the iteration.  We have deleted the current
2478                  * element and want to clear ATEDISK so the iteration does
2479                  * not skip the element after, which now becomes the current
2480                  * element.  This element must be re-tested if doing an
2481                  * iteration, which is handled by the RETEST flag.
2482                  */
2483                 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2484                         cursor->flags |= HAMMER_CURSOR_RETEST;
2485                         cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2486                 }
2487
2488                 /*
2489                  * An on-disk record cannot have the same delete_tid
2490                  * as its create_tid.  In a chain of record updates
2491                  * this could result in a duplicate record.
2492                  */
2493                 KKASSERT(elm->leaf.base.delete_tid !=
2494                          elm->leaf.base.create_tid);
2495         }
2496
2497         /*
2498          * Destroy the B-Tree element if asked (typically if a nohistory
2499          * file or mount, or when called by the pruning code).
2500          *
2501          * Adjust the ATEDISK flag to properly support iterations.
2502          */
2503         if (delete_flags & HAMMER_DELETE_DESTROY) {
2504                 data_offset = elm->leaf.data_offset;
2505                 data_len = elm->leaf.data_len;
2506                 rec_type = elm->leaf.base.rec_type;
2507                 if (doprop) {
2508                         save_leaf = elm->leaf;
2509                         leaf = &save_leaf;
2510                 }
2511                 if (elm->base.rec_type == HAMMER_RECTYPE_INODE &&
2512                     elm->leaf.base.delete_tid == 0) {
2513                         icount = -1;
2514                 }
2515
2516                 error = hammer_btree_delete(cursor);
2517                 if (error == 0) {
2518                         /*
2519                          * The deletion moves the next element (if any) to
2520                          * the current element position.  We must clear
2521                          * ATEDISK so this element is not skipped and we
2522                          * must set RETEST to force any iteration to re-test
2523                          * the element.
2524                          */
2525                         if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) {
2526                                 cursor->flags |= HAMMER_CURSOR_RETEST;
2527                                 cursor->flags &= ~HAMMER_CURSOR_ATEDISK;
2528                         }
2529                 }
2530                 if (error == 0) {
2531                         switch(data_offset & HAMMER_OFF_ZONE_MASK) {
2532                         case HAMMER_ZONE_LARGE_DATA:
2533                         case HAMMER_ZONE_SMALL_DATA:
2534                         case HAMMER_ZONE_META:
2535                                 hammer_blockmap_free(trans,
2536                                                      data_offset, data_len);
2537                                 break;
2538                         default:
2539                                 break;
2540                         }
2541                 }
2542         }
2543
2544         /*
2545          * Track inode count and next_tid.  This is used by the mirroring
2546          * and PFS code.  icount can be negative, zero, or positive.
2547          */
2548         if (error == 0 && track) {
2549                 if (icount) {
2550                         hammer_modify_volume_field(trans, trans->rootvol,
2551                                                    vol0_stat_inodes);
2552                         trans->rootvol->ondisk->vol0_stat_inodes += icount;
2553                         hammer_modify_volume_done(trans->rootvol);
2554                 }
2555                 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) {
2556                         hammer_modify_volume(trans, trans->rootvol, NULL, 0);
2557                         trans->rootvol->ondisk->vol0_next_tid = delete_tid;
2558                         hammer_modify_volume_done(trans->rootvol);
2559                 }
2560         }
2561
2562         /*
2563          * mirror_tid propagation occurs if the node's mirror_tid had to be
2564          * updated while adjusting the delete_tid.
2565          *
2566          * This occurs when deleting even in nohistory mode, but does not
2567          * occur when pruning an already-deleted node.
2568          *
2569          * cursor->ip is NULL when called from the pruning, mirroring,
2570          * and pfs code.  If non-NULL propagation will be conditionalized
2571          * on whether the PFS is in no-history mode or not.
2572          *
2573          * WARNING: cursor's leaf pointer may have changed after do_propagation
2574          *          returns!
2575          */
2576         if (doprop) {
2577                 if (cursor->ip)
2578                         hammer_btree_do_propagation(cursor, cursor->ip->pfsm, leaf);
2579                 else
2580                         hammer_btree_do_propagation(cursor, NULL, leaf);
2581         }
2582         hammer_sync_unlock(trans);
2583         return (error);
2584 }
2585
2586 /*
2587  * Determine whether we can remove a directory.  This routine checks whether
2588  * a directory is empty or not and enforces flush connectivity.
2589  *
2590  * Flush connectivity requires that we block if the target directory is
2591  * currently flushing, otherwise it may not end up in the same flush group.
2592  *
2593  * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure.
2594  */
2595 int
2596 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip)
2597 {
2598         struct hammer_cursor cursor;
2599         int error;
2600
2601         /*
2602          * Check directory empty
2603          */
2604         hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
2605
2606         cursor.key_beg.localization = ip->obj_localization +
2607                                       hammer_dir_localization(ip);
2608         cursor.key_beg.obj_id = ip->obj_id;
2609         cursor.key_beg.create_tid = 0;
2610         cursor.key_beg.delete_tid = 0;
2611         cursor.key_beg.obj_type = 0;
2612         cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE + 1;
2613         cursor.key_beg.key = HAMMER_MIN_KEY;
2614
2615         cursor.key_end = cursor.key_beg;
2616         cursor.key_end.rec_type = 0xFFFF;
2617         cursor.key_end.key = HAMMER_MAX_KEY;
2618
2619         cursor.asof = ip->obj_asof;
2620         cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
2621
2622         error = hammer_ip_first(&cursor);
2623         if (error == ENOENT)
2624                 error = 0;
2625         else if (error == 0)
2626                 error = ENOTEMPTY;
2627         hammer_done_cursor(&cursor);
2628         return(error);
2629 }
2630
2631 /*
2632  * Localize the data payload.  Directory entries may need their
2633  * localization adjusted.
2634  */
2635 static
2636 int
2637 hammer_cursor_localize_data(hammer_data_ondisk_t data,
2638                             hammer_btree_leaf_elm_t leaf)
2639 {
2640         u_int32_t localization;
2641
2642         if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) {
2643                 localization = leaf->base.localization &
2644                                HAMMER_LOCALIZE_PSEUDOFS_MASK;
2645                 if (data->entry.localization != localization) {
2646                         data->entry.localization = localization;
2647                         hammer_crc_set_leaf(data, leaf);
2648                 }
2649         }
2650         return(0);
2651 }