1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include "hammer.h" 36 37 static int hammer_mem_lookup(hammer_cursor_t cursor); 38 static int hammer_mem_first(hammer_cursor_t cursor); 39 static int hammer_frontend_trunc_callback(hammer_record_t record, 40 void *data __unused); 41 static int hammer_bulk_scan_callback(hammer_record_t record, void *data); 42 static int hammer_record_needs_overwrite_delete(hammer_record_t record); 43 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip, 44 hammer_btree_leaf_elm_t leaf); 45 static int hammer_cursor_localize_data(hammer_mount_t hmp, hammer_data_ondisk_t data, 46 hammer_btree_leaf_elm_t leaf); 47 48 struct rec_trunc_info { 49 uint16_t rec_type; 50 int64_t trunc_off; 51 }; 52 53 struct hammer_bulk_info { 54 hammer_record_t record; 55 hammer_record_t conflict; 56 }; 57 58 /* 59 * Red-black tree support. Comparison code for insertion. 60 */ 61 static int 62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2) 63 { 64 if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type) 65 return(-1); 66 if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type) 67 return(1); 68 69 if (rec1->leaf.base.key < rec2->leaf.base.key) 70 return(-1); 71 if (rec1->leaf.base.key > rec2->leaf.base.key) 72 return(1); 73 74 /* 75 * For search & insertion purposes records deleted by the 76 * frontend or deleted/committed by the backend are silently 77 * ignored. Otherwise pipelined insertions will get messed 78 * up. 79 * 80 * rec1 is greater then rec2 if rec1 is marked deleted. 81 * rec1 is less then rec2 if rec2 is marked deleted. 82 * 83 * Multiple deleted records may be present, do not return 0 84 * if both are marked deleted. 85 */ 86 if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 87 HAMMER_RECF_COMMITTED)) { 88 return(1); 89 } 90 if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 91 HAMMER_RECF_COMMITTED)) { 92 return(-1); 93 } 94 95 return(0); 96 } 97 98 /* 99 * Basic record comparison code similar to hammer_btree_cmp(). 100 * 101 * obj_id is not compared and may not yet be assigned in the record. 102 */ 103 static int 104 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec) 105 { 106 if (elm->rec_type < rec->leaf.base.rec_type) 107 return(-3); 108 if (elm->rec_type > rec->leaf.base.rec_type) 109 return(3); 110 111 if (elm->key < rec->leaf.base.key) 112 return(-2); 113 if (elm->key > rec->leaf.base.key) 114 return(2); 115 116 /* 117 * Never match against an item deleted by the frontend 118 * or backend, or committed by the backend. 119 * 120 * elm is less then rec if rec is marked deleted. 121 */ 122 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 123 HAMMER_RECF_COMMITTED)) { 124 return(-1); 125 } 126 return(0); 127 } 128 129 /* 130 * Ranged scan to locate overlapping record(s). This is used by 131 * hammer_ip_get_bulk() to locate an overlapping record. We have 132 * to use a ranged scan because the keys for data records with the 133 * same file base offset can be different due to differing data_len's. 134 * 135 * NOTE: The base file offset of a data record is (key - data_len), not (key). 136 */ 137 static int 138 hammer_rec_overlap_cmp(hammer_record_t rec, void *data) 139 { 140 struct hammer_bulk_info *info = data; 141 hammer_btree_leaf_elm_t leaf = &info->record->leaf; 142 143 if (rec->leaf.base.rec_type < leaf->base.rec_type) 144 return(-3); 145 if (rec->leaf.base.rec_type > leaf->base.rec_type) 146 return(3); 147 148 /* 149 * Overlap compare 150 */ 151 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) { 152 /* rec_beg >= leaf_end */ 153 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key) 154 return(2); 155 /* rec_end <= leaf_beg */ 156 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len) 157 return(-2); 158 } else { 159 if (rec->leaf.base.key < leaf->base.key) 160 return(-2); 161 if (rec->leaf.base.key > leaf->base.key) 162 return(2); 163 } 164 165 /* 166 * We have to return 0 at this point, even if DELETED_FE is set, 167 * because returning anything else will cause the scan to ignore 168 * one of the branches when we really want it to check both. 169 */ 170 return(0); 171 } 172 173 /* 174 * RB_SCAN comparison code for hammer_mem_first(). The argument order 175 * is reversed so the comparison result has to be negated. key_beg and 176 * key_end are both range-inclusive. 177 * 178 * Localized deletions are not cached in-memory. 179 */ 180 static 181 int 182 hammer_rec_scan_cmp(hammer_record_t rec, void *data) 183 { 184 hammer_cursor_t cursor = data; 185 int r; 186 187 r = hammer_rec_cmp(&cursor->key_beg, rec); 188 if (r > 1) 189 return(-1); 190 r = hammer_rec_cmp(&cursor->key_end, rec); 191 if (r < -1) 192 return(1); 193 return(0); 194 } 195 196 /* 197 * This compare function is used when simply looking up key_beg. 198 */ 199 static 200 int 201 hammer_rec_find_cmp(hammer_record_t rec, void *data) 202 { 203 hammer_cursor_t cursor = data; 204 int r; 205 206 r = hammer_rec_cmp(&cursor->key_beg, rec); 207 if (r > 1) 208 return(-1); 209 if (r < -1) 210 return(1); 211 return(0); 212 } 213 214 /* 215 * Locate blocks within the truncation range. Partial blocks do not count. 216 */ 217 static 218 int 219 hammer_rec_trunc_cmp(hammer_record_t rec, void *data) 220 { 221 struct rec_trunc_info *info = data; 222 223 if (rec->leaf.base.rec_type < info->rec_type) 224 return(-1); 225 if (rec->leaf.base.rec_type > info->rec_type) 226 return(1); 227 228 switch(rec->leaf.base.rec_type) { 229 case HAMMER_RECTYPE_DB: 230 /* 231 * DB record key is not beyond the truncation point, retain. 232 */ 233 if (rec->leaf.base.key < info->trunc_off) 234 return(-1); 235 break; 236 case HAMMER_RECTYPE_DATA: 237 /* 238 * DATA record offset start is not beyond the truncation point, 239 * retain. 240 */ 241 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off) 242 return(-1); 243 break; 244 default: 245 hpanic("unexpected record type"); 246 } 247 248 /* 249 * The record start is >= the truncation point, return match, 250 * the record should be destroyed. 251 */ 252 return(0); 253 } 254 255 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare); 256 257 /* 258 * Allocate a record for the caller to finish filling in. The record is 259 * returned referenced. In order to manually set data call this function 260 * with data_len=0 and then manually set record->leaf.data_len and 261 * record->data later. 262 */ 263 hammer_record_t 264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len) 265 { 266 hammer_record_t record; 267 hammer_mount_t hmp; 268 269 hmp = ip->hmp; 270 ++hammer_count_records; 271 record = kmalloc(sizeof(*record), hmp->m_misc, 272 M_WAITOK | M_ZERO | M_USE_RESERVE); 273 record->flush_state = HAMMER_FST_IDLE; 274 record->ip = ip; 275 record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD; 276 record->leaf.data_len = data_len; 277 hammer_ref(&record->lock); 278 279 if (data_len) { 280 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO); 281 record->flags |= HAMMER_RECF_ALLOCDATA; 282 ++hammer_count_record_datas; 283 } 284 285 return (record); 286 } 287 288 void 289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident) 290 { 291 while (record->flush_state == HAMMER_FST_FLUSH) { 292 record->flags |= HAMMER_RECF_WANTED; 293 tsleep(record, 0, ident, 0); 294 } 295 } 296 297 /* 298 * Called from the backend, hammer_inode.c, after a record has been 299 * flushed to disk. The record has been exclusively locked by the 300 * caller and interlocked with BE. 301 * 302 * We clean up the state, unlock, and release the record (the record 303 * was referenced by the fact that it was in the HAMMER_FST_FLUSH state). 304 */ 305 void 306 hammer_flush_record_done(hammer_record_t record, int error) 307 { 308 hammer_inode_t target_ip; 309 310 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 311 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE); 312 313 /* 314 * If an error occured, the backend was unable to sync the 315 * record to its media. Leave the record intact. 316 */ 317 if (error) { 318 hammer_critical_error(record->ip->hmp, record->ip, error, 319 "while flushing record"); 320 } 321 322 --record->flush_group->refs; 323 record->flush_group = NULL; 324 325 /* 326 * Adjust the flush state and dependancy based on success or 327 * failure. 328 */ 329 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) { 330 if ((target_ip = record->target_ip) != NULL) { 331 TAILQ_REMOVE(&target_ip->target_list, record, 332 target_entry); 333 record->target_ip = NULL; 334 hammer_test_inode(target_ip); 335 } 336 record->flush_state = HAMMER_FST_IDLE; 337 } else { 338 if (record->target_ip) { 339 record->flush_state = HAMMER_FST_SETUP; 340 hammer_test_inode(record->ip); 341 hammer_test_inode(record->target_ip); 342 } else { 343 record->flush_state = HAMMER_FST_IDLE; 344 } 345 } 346 record->flags &= ~HAMMER_RECF_INTERLOCK_BE; 347 348 /* 349 * Cleanup 350 */ 351 if (record->flags & HAMMER_RECF_WANTED) { 352 record->flags &= ~HAMMER_RECF_WANTED; 353 wakeup(record); 354 } 355 hammer_rel_mem_record(record); 356 } 357 358 /* 359 * Release a memory record. Records marked for deletion are immediately 360 * removed from the RB-Tree but otherwise left intact until the last ref 361 * goes away. 362 */ 363 void 364 hammer_rel_mem_record(hammer_record_t record) 365 { 366 hammer_mount_t hmp; 367 hammer_reserve_t resv; 368 hammer_inode_t ip; 369 hammer_inode_t target_ip; 370 int diddrop; 371 372 hammer_rel(&record->lock); 373 374 if (hammer_norefs(&record->lock)) { 375 /* 376 * Upon release of the last reference wakeup any waiters. 377 * The record structure may get destroyed so callers will 378 * loop up and do a relookup. 379 * 380 * WARNING! Record must be removed from RB-TREE before we 381 * might possibly block. hammer_test_inode() can block! 382 */ 383 ip = record->ip; 384 hmp = ip->hmp; 385 386 /* 387 * Upon release of the last reference a record marked deleted 388 * by the front or backend, or committed by the backend, 389 * is destroyed. 390 */ 391 if (record->flags & (HAMMER_RECF_DELETED_FE | 392 HAMMER_RECF_DELETED_BE | 393 HAMMER_RECF_COMMITTED)) { 394 KKASSERT(hammer_isactive(&ip->lock) > 0); 395 KKASSERT(record->flush_state != HAMMER_FST_FLUSH); 396 397 /* 398 * target_ip may have zero refs, we have to ref it 399 * to prevent it from being ripped out from under 400 * us. 401 */ 402 if ((target_ip = record->target_ip) != NULL) { 403 TAILQ_REMOVE(&target_ip->target_list, 404 record, target_entry); 405 record->target_ip = NULL; 406 hammer_ref(&target_ip->lock); 407 } 408 409 /* 410 * Remove the record from the RB-Tree 411 */ 412 if (record->flags & HAMMER_RECF_ONRBTREE) { 413 RB_REMOVE(hammer_rec_rb_tree, 414 &ip->rec_tree, 415 record); 416 record->flags &= ~HAMMER_RECF_ONRBTREE; 417 KKASSERT(ip->rsv_recs > 0); 418 if (RB_EMPTY(&ip->rec_tree)) { 419 ip->flags &= ~HAMMER_INODE_XDIRTY; 420 ip->sync_flags &= ~HAMMER_INODE_XDIRTY; 421 } 422 diddrop = 1; 423 } else { 424 diddrop = 0; 425 } 426 427 /* 428 * We must wait for any direct-IO to complete before 429 * we can destroy the record because the bio may 430 * have a reference to it. 431 */ 432 if (record->gflags & 433 (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) { 434 hammer_io_direct_wait(record); 435 } 436 437 /* 438 * Account for the completion after the direct IO 439 * has completed. 440 */ 441 if (diddrop) { 442 --hmp->rsv_recs; 443 --ip->rsv_recs; 444 hmp->rsv_databytes -= record->leaf.data_len; 445 446 if (RB_EMPTY(&ip->rec_tree)) 447 hammer_test_inode(ip); 448 if ((ip->flags & HAMMER_INODE_RECSW) && 449 ip->rsv_recs <= hammer_limit_inode_recs/2) { 450 ip->flags &= ~HAMMER_INODE_RECSW; 451 wakeup(&ip->rsv_recs); 452 } 453 } 454 455 /* 456 * Do this test after removing record from the RB-Tree. 457 */ 458 if (target_ip) { 459 hammer_test_inode(target_ip); 460 hammer_rel_inode(target_ip, 0); 461 } 462 463 if (record->flags & HAMMER_RECF_ALLOCDATA) { 464 --hammer_count_record_datas; 465 kfree(record->data, hmp->m_misc); 466 record->flags &= ~HAMMER_RECF_ALLOCDATA; 467 } 468 469 /* 470 * Release the reservation. 471 * 472 * If the record was not committed we can theoretically 473 * undo the reservation. However, doing so might 474 * create weird edge cases with the ordering of 475 * direct writes because the related buffer cache 476 * elements are per-vnode. So we don't try. 477 */ 478 if ((resv = record->resv) != NULL) { 479 /* XXX undo leaf.data_offset,leaf.data_len */ 480 hammer_blockmap_reserve_complete(hmp, resv); 481 record->resv = NULL; 482 } 483 record->data = NULL; 484 --hammer_count_records; 485 kfree(record, hmp->m_misc); 486 } 487 } 488 } 489 490 /* 491 * Record visibility depends on whether the record is being accessed by 492 * the backend or the frontend. Backend tests ignore the frontend delete 493 * flag. Frontend tests do NOT ignore the backend delete/commit flags and 494 * must also check for commit races. 495 * 496 * Return non-zero if the record is visible, zero if it isn't or if it is 497 * deleted. Returns 0 if the record has been comitted (unless the special 498 * delete-visibility flag is set). A committed record must be located 499 * via the media B-Tree. Returns non-zero if the record is good. 500 * 501 * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory 502 * records to be returned. This is so pending deletions are detected 503 * when using an iterator to locate an unused hash key, or when we need 504 * to locate historical records on-disk to destroy. 505 */ 506 static __inline 507 int 508 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record) 509 { 510 if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) 511 return(1); 512 if (cursor->flags & HAMMER_CURSOR_BACKEND) { 513 if (record->flags & (HAMMER_RECF_DELETED_BE | 514 HAMMER_RECF_COMMITTED)) { 515 return(0); 516 } 517 } else { 518 if (record->flags & (HAMMER_RECF_DELETED_FE | 519 HAMMER_RECF_DELETED_BE | 520 HAMMER_RECF_COMMITTED)) { 521 return(0); 522 } 523 } 524 return(1); 525 } 526 527 /* 528 * This callback is used as part of the RB_SCAN function for in-memory 529 * records. We terminate it (return -1) as soon as we get a match. 530 * 531 * This routine is used by frontend code. 532 * 533 * The primary compare code does not account for ASOF lookups. This 534 * code handles that case as well as a few others. 535 */ 536 static 537 int 538 hammer_rec_scan_callback(hammer_record_t rec, void *data) 539 { 540 hammer_cursor_t cursor = data; 541 542 /* 543 * We terminate on success, so this should be NULL on entry. 544 */ 545 KKASSERT(cursor->iprec == NULL); 546 547 /* 548 * Skip if the record was marked deleted or committed. 549 */ 550 if (hammer_ip_iterate_mem_good(cursor, rec) == 0) 551 return(0); 552 553 /* 554 * Skip if not visible due to our as-of TID 555 */ 556 if (cursor->flags & HAMMER_CURSOR_ASOF) { 557 if (cursor->asof < rec->leaf.base.create_tid) 558 return(0); 559 if (rec->leaf.base.delete_tid && 560 cursor->asof >= rec->leaf.base.delete_tid) { 561 return(0); 562 } 563 } 564 565 /* 566 * ref the record. The record is protected from backend B-Tree 567 * interactions by virtue of the cursor's IP lock. 568 */ 569 hammer_ref(&rec->lock); 570 571 /* 572 * The record may have been deleted or committed while we 573 * were blocked. XXX remove? 574 */ 575 if (hammer_ip_iterate_mem_good(cursor, rec) == 0) { 576 hammer_rel_mem_record(rec); 577 return(0); 578 } 579 580 /* 581 * Set the matching record and stop the scan. 582 */ 583 cursor->iprec = rec; 584 return(-1); 585 } 586 587 588 /* 589 * Lookup an in-memory record given the key specified in the cursor. Works 590 * just like hammer_btree_lookup() but operates on an inode's in-memory 591 * record list. 592 * 593 * The lookup must fail if the record is marked for deferred deletion. 594 * 595 * The API for mem/btree_lookup() does not mess with the ATE/EOF bits. 596 */ 597 static 598 int 599 hammer_mem_lookup(hammer_cursor_t cursor) 600 { 601 KKASSERT(cursor->ip != NULL); 602 if (cursor->iprec) { 603 hammer_rel_mem_record(cursor->iprec); 604 cursor->iprec = NULL; 605 } 606 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp, 607 hammer_rec_scan_callback, cursor); 608 609 return (cursor->iprec ? 0 : ENOENT); 610 } 611 612 /* 613 * hammer_mem_first() - locate the first in-memory record matching the 614 * cursor within the bounds of the key range. 615 * 616 * WARNING! API is slightly different from btree_first(). hammer_mem_first() 617 * will set ATEMEM the same as MEMEOF, and does not return any error. 618 */ 619 static 620 int 621 hammer_mem_first(hammer_cursor_t cursor) 622 { 623 KKASSERT(cursor->ip != NULL); 624 if (cursor->iprec) { 625 hammer_rel_mem_record(cursor->iprec); 626 cursor->iprec = NULL; 627 } 628 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_scan_cmp, 629 hammer_rec_scan_callback, cursor); 630 631 if (cursor->iprec) 632 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM); 633 else 634 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM; 635 636 return (cursor->iprec ? 0 : ENOENT); 637 } 638 639 /************************************************************************ 640 * HAMMER IN-MEMORY RECORD FUNCTIONS * 641 ************************************************************************ 642 * 643 * These functions manipulate in-memory records. Such records typically 644 * exist prior to being committed to disk or indexed via the on-disk B-Tree. 645 */ 646 647 /* 648 * Add a directory entry (dip,ncp) which references inode (ip). 649 * 650 * Note that the low 32 bits of the namekey are set temporarily to create 651 * a unique in-memory record, and may be modified a second time when the 652 * record is synchronized to disk. In particular, the low 32 bits cannot be 653 * all 0's when synching to disk, which is not handled here. 654 * 655 * NOTE: bytes does not include any terminating \0 on name, and name might 656 * not be terminated. 657 */ 658 int 659 hammer_ip_add_direntry(hammer_transaction_t trans, 660 hammer_inode_t dip, const char *name, int bytes, 661 hammer_inode_t ip) 662 { 663 struct hammer_cursor cursor; 664 hammer_record_t record; 665 int error; 666 uint32_t max_iterations; 667 668 KKASSERT(dip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY); 669 670 record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes)); 671 672 record->type = HAMMER_MEM_RECORD_ADD; 673 record->leaf.base.localization = dip->obj_localization | 674 hammer_dir_localization(dip); 675 record->leaf.base.obj_id = dip->obj_id; 676 record->leaf.base.key = hammer_direntry_namekey(dip, name, bytes, 677 &max_iterations); 678 record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY; 679 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 680 record->data->entry.obj_id = ip->obj_id; 681 record->data->entry.localization = ip->obj_localization; 682 bcopy(name, record->data->entry.name, bytes); 683 684 ++ip->ino_data.nlinks; 685 ip->ino_data.ctime = trans->time; 686 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY); 687 688 /* 689 * Find an unused namekey. Both the in-memory record tree and 690 * the B-Tree are checked. We do not want historically deleted 691 * names to create a collision as our iteration space may be limited, 692 * and since create_tid wouldn't match anyway an ASOF search 693 * must be used to locate collisions. 694 * 695 * delete-visibility is set so pending deletions do not give us 696 * a false-negative on our ability to use an iterator. 697 * 698 * The iterator must not rollover the key. Directory keys only 699 * use the positive key space. 700 */ 701 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip); 702 cursor.key_beg = record->leaf.base; 703 cursor.flags |= HAMMER_CURSOR_ASOF; 704 cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 705 cursor.asof = ip->obj_asof; 706 707 while (hammer_ip_lookup(&cursor) == 0) { 708 ++record->leaf.base.key; 709 KKASSERT(record->leaf.base.key > 0); 710 cursor.key_beg.key = record->leaf.base.key; 711 if (--max_iterations == 0) { 712 hammer_rel_mem_record(record); 713 hmkprintf(trans->hmp, "Failed to find an unused namekey\n"); 714 error = ENOSPC; 715 goto failed; 716 } 717 } 718 719 /* 720 * The target inode and the directory entry are bound together. 721 */ 722 record->target_ip = ip; 723 record->flush_state = HAMMER_FST_SETUP; 724 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry); 725 726 /* 727 * The inode now has a dependancy and must be taken out of the idle 728 * state. An inode not in an idle state is given an extra reference. 729 * 730 * When transitioning to a SETUP state flag for an automatic reflush 731 * when the dependancies are disposed of if someone is waiting on 732 * the inode. 733 */ 734 if (ip->flush_state == HAMMER_FST_IDLE) { 735 hammer_ref(&ip->lock); 736 ip->flush_state = HAMMER_FST_SETUP; 737 if (ip->flags & HAMMER_INODE_FLUSHW) 738 ip->flags |= HAMMER_INODE_REFLUSH; 739 } 740 error = hammer_mem_add(record); 741 if (error == 0) { 742 dip->ino_data.mtime = trans->time; 743 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME); 744 } 745 failed: 746 hammer_done_cursor(&cursor); 747 return(error); 748 } 749 750 /* 751 * Delete the directory entry and update the inode link count. The 752 * cursor must be seeked to the directory entry record being deleted. 753 * 754 * The related inode should be share-locked by the caller. The caller is 755 * on the frontend. It could also be NULL indicating that the directory 756 * entry being removed has no related inode. 757 * 758 * This function can return EDEADLK requiring the caller to terminate 759 * the cursor, any locks, wait on the returned record, and retry. 760 */ 761 int 762 hammer_ip_del_direntry(hammer_transaction_t trans, 763 hammer_cursor_t cursor, hammer_inode_t dip, 764 hammer_inode_t ip) 765 { 766 hammer_record_t record; 767 int error; 768 769 if (hammer_cursor_inmem(cursor)) { 770 /* 771 * In-memory (unsynchronized) records can simply be freed. 772 * 773 * Even though the HAMMER_RECF_DELETED_FE flag is ignored 774 * by the backend, we must still avoid races against the 775 * backend potentially syncing the record to the media. 776 * 777 * We cannot call hammer_ip_delete_record(), that routine may 778 * only be called from the backend. 779 */ 780 record = cursor->iprec; 781 if (record->flags & (HAMMER_RECF_INTERLOCK_BE | 782 HAMMER_RECF_DELETED_BE | 783 HAMMER_RECF_COMMITTED)) { 784 KKASSERT(cursor->deadlk_rec == NULL); 785 hammer_ref(&record->lock); 786 cursor->deadlk_rec = record; 787 error = EDEADLK; 788 } else { 789 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD); 790 record->flags |= HAMMER_RECF_DELETED_FE; 791 error = 0; 792 } 793 } else { 794 /* 795 * If the record is on-disk we have to queue the deletion by 796 * the record's key. This also causes lookups to skip the 797 * record (lookups for the purposes of finding an unused 798 * directory key do not skip the record). 799 */ 800 KKASSERT(dip->flags & 801 (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK)); 802 record = hammer_alloc_mem_record(dip, 0); 803 record->type = HAMMER_MEM_RECORD_DEL; 804 record->leaf.base = cursor->leaf->base; 805 KKASSERT(dip->obj_id == record->leaf.base.obj_id); 806 807 /* 808 * ip may be NULL, indicating the deletion of a directory 809 * entry which has no related inode. 810 */ 811 record->target_ip = ip; 812 if (ip) { 813 record->flush_state = HAMMER_FST_SETUP; 814 TAILQ_INSERT_TAIL(&ip->target_list, record, 815 target_entry); 816 } else { 817 record->flush_state = HAMMER_FST_IDLE; 818 } 819 820 /* 821 * The inode now has a dependancy and must be taken out of 822 * the idle state. An inode not in an idle state is given 823 * an extra reference. 824 * 825 * When transitioning to a SETUP state flag for an automatic 826 * reflush when the dependancies are disposed of if someone 827 * is waiting on the inode. 828 */ 829 if (ip && ip->flush_state == HAMMER_FST_IDLE) { 830 hammer_ref(&ip->lock); 831 ip->flush_state = HAMMER_FST_SETUP; 832 if (ip->flags & HAMMER_INODE_FLUSHW) 833 ip->flags |= HAMMER_INODE_REFLUSH; 834 } 835 836 error = hammer_mem_add(record); 837 } 838 839 /* 840 * One less link. The file may still be open in the OS even after 841 * all links have gone away. 842 * 843 * We have to terminate the cursor before syncing the inode to 844 * avoid deadlocking against ourselves. XXX this may no longer 845 * be true. 846 * 847 * If nlinks drops to zero and the vnode is inactive (or there is 848 * no vnode), call hammer_inode_unloadable_check() to zonk the 849 * inode. If we don't do this here the inode will not be destroyed 850 * on-media until we unmount. 851 */ 852 if (error == 0) { 853 if (ip) { 854 --ip->ino_data.nlinks; /* do before we might block */ 855 ip->ino_data.ctime = trans->time; 856 } 857 dip->ino_data.mtime = trans->time; 858 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME); 859 if (ip) { 860 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY); 861 if (ip->ino_data.nlinks == 0 && 862 (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) { 863 hammer_done_cursor(cursor); 864 hammer_inode_unloadable_check(ip, 1); 865 hammer_flush_inode(ip, 0); 866 } 867 } 868 869 } 870 return(error); 871 } 872 873 /* 874 * Add a record to an inode. 875 * 876 * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and 877 * initialize the following additional fields that are not initialized by these 878 * functions. 879 * 880 * The related inode should be share-locked by the caller. The caller is 881 * on the frontend. 882 * 883 * record->leaf.base.key 884 * record->leaf.base.rec_type 885 * record->leaf.base.localization 886 */ 887 int 888 hammer_ip_add_record(hammer_transaction_t trans, hammer_record_t record) 889 { 890 hammer_inode_t ip = record->ip; 891 int error; 892 893 KKASSERT(record->leaf.base.localization != 0); 894 record->leaf.base.obj_id = ip->obj_id; 895 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 896 error = hammer_mem_add(record); 897 return(error); 898 } 899 900 /* 901 * Locate a pre-existing bulk record in memory. The caller wishes to 902 * replace the record with a new one. The existing record may have a 903 * different length (and thus a different key) so we have to use an 904 * overlap check function. 905 */ 906 static hammer_record_t 907 hammer_ip_get_bulk(hammer_record_t record) 908 { 909 struct hammer_bulk_info info; 910 hammer_inode_t ip = record->ip; 911 912 info.record = record; 913 info.conflict = NULL; 914 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp, 915 hammer_bulk_scan_callback, &info); 916 917 return(info.conflict); /* may be NULL */ 918 } 919 920 /* 921 * Take records vetted by overlap_cmp. The first non-deleted record 922 * (if any) stops the scan. 923 */ 924 static int 925 hammer_bulk_scan_callback(hammer_record_t record, void *data) 926 { 927 struct hammer_bulk_info *info = data; 928 929 if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 930 HAMMER_RECF_COMMITTED)) { 931 return(0); 932 } 933 hammer_ref(&record->lock); 934 info->conflict = record; 935 return(-1); /* stop scan */ 936 } 937 938 /* 939 * Reserve blockmap space placemarked with an in-memory record. 940 * 941 * This routine is called by the frontend in order to be able to directly 942 * flush a buffer cache buffer. The frontend has locked the related buffer 943 * cache buffers and we should be able to manipulate any overlapping 944 * in-memory records. 945 * 946 * The caller is responsible for adding the returned record and deleting 947 * the returned conflicting record (if any), typically by calling 948 * hammer_ip_replace_bulk() (via hammer_io_direct_write()). 949 */ 950 hammer_record_t 951 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes, 952 int *errorp) 953 { 954 hammer_record_t record; 955 int zone; 956 957 /* 958 * Create a record to cover the direct write. The record cannot 959 * be added to the in-memory RB tree here as it might conflict 960 * with an existing memory record. See hammer_io_direct_write(). 961 * 962 * The backend is responsible for finalizing the space reserved in 963 * this record. 964 * 965 * XXX bytes not aligned, depend on the reservation code to 966 * align the reservation. 967 */ 968 record = hammer_alloc_mem_record(ip, 0); 969 zone = hammer_data_zone_index(bytes); 970 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes, 971 &record->leaf.data_offset, 972 errorp); 973 if (record->resv == NULL) { 974 hdkprintf("reservation failed\n"); 975 hammer_rel_mem_record(record); 976 return(NULL); 977 } 978 record->type = HAMMER_MEM_RECORD_DATA; 979 record->leaf.base.rec_type = HAMMER_RECTYPE_DATA; 980 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 981 record->leaf.base.obj_id = ip->obj_id; 982 record->leaf.base.key = file_offset + bytes; 983 record->leaf.base.localization = ip->obj_localization | 984 HAMMER_LOCALIZE_MISC; 985 record->leaf.data_len = bytes; 986 hammer_crc_set_leaf(ip->hmp->version, data, &record->leaf); 987 KKASSERT(*errorp == 0); 988 989 return(record); 990 } 991 992 /* 993 * Called by hammer_io_direct_write() prior to any possible completion 994 * of the BIO to emplace the memory record associated with the I/O and 995 * to replace any prior memory record which might still be active. 996 * 997 * Setting the FE deleted flag on the old record (if any) avoids any RB 998 * tree insertion conflict, amoung other things. 999 * 1000 * This has to be done prior to the caller completing any related buffer 1001 * cache I/O or a reinstantiation of the buffer may load data from the 1002 * old media location instead of the new media location. The holding 1003 * of the locked buffer cache buffer serves to interlock the record 1004 * replacement operation. 1005 */ 1006 void 1007 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record) 1008 { 1009 hammer_record_t conflict; 1010 int error __debugvar; 1011 1012 while ((conflict = hammer_ip_get_bulk(record)) != NULL) { 1013 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) { 1014 conflict->flags |= HAMMER_RECF_DELETED_FE; 1015 break; 1016 } 1017 conflict->flags |= HAMMER_RECF_WANTED; 1018 tsleep(conflict, 0, "hmrrc3", 0); 1019 hammer_rel_mem_record(conflict); 1020 } 1021 error = hammer_mem_add(record); 1022 if (conflict) 1023 hammer_rel_mem_record(conflict); 1024 KKASSERT(error == 0); 1025 } 1026 1027 /* 1028 * Frontend truncation code. Scan in-memory records only. On-disk records 1029 * and records in a flushing state are handled by the backend. The vnops 1030 * setattr code will handle the block containing the truncation point. 1031 * 1032 * Partial blocks are not deleted. 1033 * 1034 * This code is only called on regular files. 1035 */ 1036 int 1037 hammer_ip_frontend_trunc(hammer_inode_t ip, off_t file_size) 1038 { 1039 struct rec_trunc_info info; 1040 1041 switch(ip->ino_data.obj_type) { 1042 case HAMMER_OBJTYPE_REGFILE: 1043 info.rec_type = HAMMER_RECTYPE_DATA; 1044 break; 1045 case HAMMER_OBJTYPE_DBFILE: 1046 info.rec_type = HAMMER_RECTYPE_DB; 1047 break; 1048 default: 1049 return(EINVAL); 1050 } 1051 info.trunc_off = file_size; 1052 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp, 1053 hammer_frontend_trunc_callback, &info); 1054 return(0); 1055 } 1056 1057 /* 1058 * Scan callback for frontend records to destroy during a truncation. 1059 * We must ensure that DELETED_FE is set on the record or the frontend 1060 * will get confused in future read() calls. 1061 * 1062 * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held. 1063 * In this rare case we must wait for the interlock to be cleared. 1064 * 1065 * NOTE: This function is only called on regular files. There are further 1066 * restrictions to the setting of DELETED_FE on directory records 1067 * undergoing a flush due to sensitive inode link count calculations. 1068 */ 1069 static int 1070 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused) 1071 { 1072 if (record->flags & HAMMER_RECF_DELETED_FE) 1073 return(0); 1074 #if 0 1075 if (record->flush_state == HAMMER_FST_FLUSH) 1076 return(0); 1077 #endif 1078 hammer_ref(&record->lock); 1079 while (record->flags & HAMMER_RECF_INTERLOCK_BE) 1080 hammer_wait_mem_record_ident(record, "hmmtrr"); 1081 record->flags |= HAMMER_RECF_DELETED_FE; 1082 hammer_rel_mem_record(record); 1083 return(0); 1084 } 1085 1086 /* 1087 * Return 1 if the caller must check for and delete existing records 1088 * before writing out a new data record. 1089 * 1090 * Return 0 if the caller can just insert the record into the B-Tree without 1091 * checking. 1092 */ 1093 static int 1094 hammer_record_needs_overwrite_delete(hammer_record_t record) 1095 { 1096 hammer_inode_t ip = record->ip; 1097 int64_t file_offset; 1098 int r; 1099 1100 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) 1101 file_offset = record->leaf.base.key; 1102 else 1103 file_offset = record->leaf.base.key - record->leaf.data_len; 1104 r = (file_offset < ip->save_trunc_off); 1105 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1106 if (ip->save_trunc_off <= record->leaf.base.key) 1107 ip->save_trunc_off = record->leaf.base.key + 1; 1108 } else { 1109 if (ip->save_trunc_off < record->leaf.base.key) 1110 ip->save_trunc_off = record->leaf.base.key; 1111 } 1112 return(r); 1113 } 1114 1115 /* 1116 * Backend code. Sync a record to the media. 1117 */ 1118 int 1119 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record) 1120 { 1121 hammer_transaction_t trans = cursor->trans; 1122 hammer_mount_t hmp = trans->hmp; 1123 int64_t file_offset; 1124 int bytes; 1125 void *bdata; 1126 int error; 1127 int doprop; 1128 1129 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 1130 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE); 1131 KKASSERT(record->leaf.base.localization != 0); 1132 1133 /* 1134 * Any direct-write related to the record must complete before we 1135 * can sync the record to the on-disk media. 1136 */ 1137 if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) 1138 hammer_io_direct_wait(record); 1139 1140 /* 1141 * If this is a bulk-data record placemarker there may be an existing 1142 * record on-disk, indicating a data overwrite. If there is the 1143 * on-disk record must be deleted before we can insert our new record. 1144 * 1145 * We've synthesized this record and do not know what the create_tid 1146 * on-disk is, nor how much data it represents. 1147 * 1148 * Keep in mind that (key) for data records is (base_offset + len), 1149 * not (base_offset). Also, we only want to get rid of on-disk 1150 * records since we are trying to sync our in-memory record, call 1151 * hammer_ip_delete_range() with truncating set to 1 to make sure 1152 * it skips in-memory records. 1153 * 1154 * It is ok for the lookup to return ENOENT. 1155 * 1156 * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have 1157 * to call hammer_ip_delete_range() or not. This also means we must 1158 * update sync_trunc_off() as we write. 1159 */ 1160 if (record->type == HAMMER_MEM_RECORD_DATA && 1161 hammer_record_needs_overwrite_delete(record)) { 1162 file_offset = record->leaf.base.key - record->leaf.data_len; 1163 bytes = HAMMER_BUFSIZE_DOALIGN(record->leaf.data_len); 1164 KKASSERT((file_offset & HAMMER_BUFMASK) == 0); 1165 error = hammer_ip_delete_range( 1166 cursor, record->ip, 1167 file_offset, file_offset + bytes - 1, 1168 1); 1169 if (error && error != ENOENT) 1170 goto done; 1171 } 1172 1173 /* 1174 * If this is a general record there may be an on-disk version 1175 * that must be deleted before we can insert the new record. 1176 */ 1177 if (record->type == HAMMER_MEM_RECORD_GENERAL) { 1178 error = hammer_delete_general(cursor, record->ip, &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 hdkprintf("DOINSERT LOOKUP %d\n", error); 1230 if (error == 0) { 1231 hdkprintf("duplicate rec at (%016jx)\n", 1232 (intmax_t)record->leaf.base.key); 1233 if (hammer_debug_critical) 1234 Debugger("duplicate record1"); 1235 error = EIO; 1236 } 1237 1238 if (error != ENOENT) 1239 goto done_unlock; 1240 1241 /* 1242 * Allocate the record and data. The result buffers will be 1243 * marked as being modified and further calls to 1244 * hammer_modify_buffer() will result in unneeded UNDO records. 1245 * 1246 * Support zero-fill records (data == NULL and data_len != 0) 1247 */ 1248 if (record->type == HAMMER_MEM_RECORD_DATA) { 1249 /* 1250 * The data portion of a bulk-data record has already been 1251 * committed to disk, we need only adjust the layer2 1252 * statistics in the same transaction as our B-Tree insert. 1253 */ 1254 KKASSERT(record->leaf.data_offset != 0); 1255 error = hammer_blockmap_finalize(trans, 1256 record->resv, 1257 record->leaf.data_offset, 1258 record->leaf.data_len); 1259 } else if (record->data && record->leaf.data_len) { 1260 /* 1261 * Wholely cached record, with data. Allocate the data. 1262 */ 1263 bdata = hammer_alloc_data(trans, record->leaf.data_len, 1264 record->leaf.base.rec_type, 1265 &record->leaf.data_offset, 1266 &cursor->data_buffer, 1267 0, &error); 1268 if (bdata == NULL) 1269 goto done_unlock; 1270 hammer_crc_set_leaf(hmp->version, record->data, &record->leaf); 1271 hammer_modify_buffer_noundo(trans, cursor->data_buffer); 1272 bcopy(record->data, bdata, record->leaf.data_len); 1273 hammer_modify_buffer_done(cursor->data_buffer); 1274 } else { 1275 /* 1276 * Wholely cached record, without data. 1277 */ 1278 record->leaf.data_offset = 0; 1279 record->leaf.data_crc = 0; 1280 } 1281 1282 error = hammer_btree_insert(cursor, &record->leaf, &doprop); 1283 if (hammer_debug_inode && error) { 1284 hdkprintf("BTREE INSERT error %d @ %016jx:%d key %016jx\n", 1285 error, 1286 (intmax_t)cursor->node->node_offset, 1287 cursor->index, 1288 (intmax_t)record->leaf.base.key); 1289 } 1290 1291 /* 1292 * Our record is on-disk and we normally mark the in-memory version 1293 * as having been committed (and not BE-deleted). 1294 * 1295 * If the record represented a directory deletion but we had to 1296 * sync a valid directory entry to disk due to dependancies, 1297 * we must convert the record to a covering delete so the 1298 * frontend does not have visibility on the synced entry. 1299 * 1300 * WARNING: cursor's leaf pointer may have changed after do_propagation 1301 * returns! 1302 */ 1303 if (error == 0) { 1304 if (doprop) { 1305 hammer_btree_do_propagation(cursor, &record->leaf); 1306 } 1307 if (record->flags & HAMMER_RECF_CONVERT_DELETE) { 1308 /* 1309 * Must convert deleted directory entry add 1310 * to a directory entry delete. 1311 */ 1312 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD); 1313 record->flags &= ~HAMMER_RECF_DELETED_FE; 1314 record->type = HAMMER_MEM_RECORD_DEL; 1315 KKASSERT(record->ip->obj_id == record->leaf.base.obj_id); 1316 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 1317 record->flags &= ~HAMMER_RECF_CONVERT_DELETE; 1318 KKASSERT((record->flags & (HAMMER_RECF_COMMITTED | 1319 HAMMER_RECF_DELETED_BE)) == 0); 1320 /* converted record is not yet committed */ 1321 /* hammer_flush_record_done takes care of the rest */ 1322 } else { 1323 /* 1324 * Everything went fine and we are now done with 1325 * this record. 1326 */ 1327 record->flags |= HAMMER_RECF_COMMITTED; 1328 ++record->ip->rec_generation; 1329 } 1330 } else { 1331 if (record->leaf.data_offset) { 1332 hammer_blockmap_free(trans, record->leaf.data_offset, 1333 record->leaf.data_len); 1334 } 1335 } 1336 done_unlock: 1337 hammer_sync_unlock(trans); 1338 done: 1339 return(error); 1340 } 1341 1342 /* 1343 * Add the record to the inode's rec_tree. The low 32 bits of a directory 1344 * entry's key is used to deal with hash collisions in the upper 32 bits. 1345 * A unique 64 bit key is generated in-memory and may be regenerated a 1346 * second time when the directory record is flushed to the on-disk B-Tree. 1347 * 1348 * A referenced record is passed to this function. This function 1349 * eats the reference. If an error occurs the record will be deleted. 1350 * 1351 * A copy of the temporary record->data pointer provided by the caller 1352 * will be made. 1353 */ 1354 int 1355 hammer_mem_add(hammer_record_t record) 1356 { 1357 hammer_mount_t hmp = record->ip->hmp; 1358 1359 /* 1360 * Make a private copy of record->data 1361 */ 1362 if (record->data) 1363 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA); 1364 1365 /* 1366 * Insert into the RB tree. A unique key should have already 1367 * been selected if this is a directory entry. 1368 */ 1369 if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) { 1370 record->flags |= HAMMER_RECF_DELETED_FE; 1371 hammer_rel_mem_record(record); 1372 return (EEXIST); 1373 } 1374 ++hmp->rsv_recs; 1375 ++record->ip->rsv_recs; 1376 record->ip->hmp->rsv_databytes += record->leaf.data_len; 1377 record->flags |= HAMMER_RECF_ONRBTREE; 1378 hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY); 1379 hammer_rel_mem_record(record); 1380 return(0); 1381 } 1382 1383 /************************************************************************ 1384 * HAMMER INODE MERGED-RECORD FUNCTIONS * 1385 ************************************************************************ 1386 * 1387 * These functions augment the B-Tree scanning functions in hammer_btree.c 1388 * by merging in-memory records with on-disk records. 1389 */ 1390 1391 /* 1392 * Locate a particular record either in-memory or on-disk. 1393 * 1394 * NOTE: This is basically a standalone routine, hammer_ip_next() may 1395 * NOT be called to iterate results. 1396 */ 1397 int 1398 hammer_ip_lookup(hammer_cursor_t cursor) 1399 { 1400 int error; 1401 1402 /* 1403 * If the element is in-memory return it without searching the 1404 * on-disk B-Tree 1405 */ 1406 KKASSERT(cursor->ip); 1407 error = hammer_mem_lookup(cursor); 1408 if (error == 0) { 1409 cursor->leaf = &cursor->iprec->leaf; 1410 return(error); 1411 } 1412 if (error != ENOENT) 1413 return(error); 1414 1415 /* 1416 * If the inode has on-disk components search the on-disk B-Tree. 1417 */ 1418 if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0) 1419 return(error); 1420 error = hammer_btree_lookup(cursor); 1421 if (error == 0) 1422 error = hammer_btree_extract_leaf(cursor); 1423 return(error); 1424 } 1425 1426 /* 1427 * Helper for hammer_ip_first()/hammer_ip_next() 1428 * 1429 * NOTE: Both ATEDISK and DISKEOF will be set the same. This sets up 1430 * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek 1431 * state if hammer_ip_next() needs to re-seek. 1432 */ 1433 static __inline 1434 int 1435 _hammer_ip_seek_btree(hammer_cursor_t cursor) 1436 { 1437 hammer_inode_t ip = cursor->ip; 1438 int error; 1439 1440 if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) { 1441 error = hammer_btree_lookup(cursor); 1442 if (error == ENOENT || error == EDEADLK) { 1443 if (hammer_debug_general & 0x2000) { 1444 hdkprintf("error %d node %p %016jx index %d\n", 1445 error, cursor->node, 1446 (intmax_t)cursor->node->node_offset, 1447 cursor->index); 1448 } 1449 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 1450 error = hammer_btree_iterate(cursor); 1451 } 1452 if (error == 0) { 1453 cursor->flags &= ~(HAMMER_CURSOR_DISKEOF | 1454 HAMMER_CURSOR_ATEDISK); 1455 } else { 1456 cursor->flags |= HAMMER_CURSOR_DISKEOF | 1457 HAMMER_CURSOR_ATEDISK; 1458 if (error == ENOENT) 1459 error = 0; 1460 } 1461 } else { 1462 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK; 1463 error = 0; 1464 } 1465 return(error); 1466 } 1467 1468 /* 1469 * Helper for hammer_ip_next() 1470 * 1471 * The caller has determined that the media cursor is further along than the 1472 * memory cursor and must be reseeked after a generation number change. 1473 */ 1474 static 1475 int 1476 _hammer_ip_reseek(hammer_cursor_t cursor) 1477 { 1478 struct hammer_base_elm save; 1479 hammer_btree_elm_t elm; 1480 int error __debugvar; 1481 int r; 1482 int again = 0; 1483 1484 /* 1485 * Do the re-seek. 1486 */ 1487 hkprintf("Debug: re-seeked during scan @ino=%016jx\n", 1488 (intmax_t)cursor->ip->obj_id); 1489 save = cursor->key_beg; 1490 cursor->key_beg = cursor->iprec->leaf.base; 1491 error = _hammer_ip_seek_btree(cursor); 1492 KKASSERT(error == 0); 1493 cursor->key_beg = save; 1494 1495 /* 1496 * If the memory record was previous returned to 1497 * the caller and the media record matches 1498 * (-1/+1: only create_tid differs), then iterate 1499 * the media record to avoid a double result. 1500 */ 1501 if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 && 1502 (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) { 1503 elm = &cursor->node->ondisk->elms[cursor->index]; 1504 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base); 1505 if (cursor->flags & HAMMER_CURSOR_ASOF) { 1506 if (r >= -1 && r <= 1) { 1507 hkprintf("Debug: iterated after " 1508 "re-seek (asof r=%d)\n", r); 1509 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1510 again = 1; 1511 } 1512 } else { 1513 if (r == 0) { 1514 hkprintf("Debug: iterated after " 1515 "re-seek\n"); 1516 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1517 again = 1; 1518 } 1519 } 1520 } 1521 return(again); 1522 } 1523 1524 /* 1525 * Locate the first record within the cursor's key_beg/key_end range, 1526 * restricted to a particular inode. 0 is returned on success, ENOENT 1527 * if no records matched the requested range, or some other error. 1528 * 1529 * When 0 is returned hammer_ip_next() may be used to iterate additional 1530 * records within the requested range. 1531 * 1532 * This function can return EDEADLK, requiring the caller to terminate 1533 * the cursor and try again. 1534 */ 1535 1536 int 1537 hammer_ip_first(hammer_cursor_t cursor) 1538 { 1539 hammer_inode_t ip __debugvar = cursor->ip; 1540 int error; 1541 1542 KKASSERT(ip != NULL); 1543 1544 /* 1545 * Clean up fields and setup for merged scan 1546 */ 1547 cursor->flags &= ~HAMMER_CURSOR_RETEST; 1548 1549 /* 1550 * Search the in-memory record list (Red-Black tree). Unlike the 1551 * B-Tree search, mem_first checks for records in the range. 1552 * 1553 * This function will setup both ATEMEM and MEMEOF properly for 1554 * the ip iteration. ATEMEM will be set if MEMEOF is set. 1555 */ 1556 hammer_mem_first(cursor); 1557 1558 /* 1559 * Detect generation changes during blockages, including 1560 * blockages which occur on the initial btree search. 1561 */ 1562 cursor->rec_generation = cursor->ip->rec_generation; 1563 1564 /* 1565 * Initial search and result 1566 */ 1567 error = _hammer_ip_seek_btree(cursor); 1568 if (error == 0) 1569 error = hammer_ip_next(cursor); 1570 1571 return (error); 1572 } 1573 1574 /* 1575 * Retrieve the next record in a merged iteration within the bounds of the 1576 * cursor. This call may be made multiple times after the cursor has been 1577 * initially searched with hammer_ip_first(). 1578 * 1579 * There are numerous special cases in this code to deal with races between 1580 * in-memory records and on-media records. 1581 * 1582 * 0 is returned on success, ENOENT if no further records match the 1583 * requested range, or some other error code is returned. 1584 */ 1585 int 1586 hammer_ip_next(hammer_cursor_t cursor) 1587 { 1588 hammer_btree_elm_t elm; 1589 hammer_record_t rec; 1590 hammer_record_t tmprec; 1591 int error; 1592 int r; 1593 1594 again: 1595 /* 1596 * Get the next on-disk record 1597 * 1598 * NOTE: If we deleted the last on-disk record we had scanned 1599 * ATEDISK will be clear and RETEST will be set, forcing 1600 * a call to iterate. The fact that ATEDISK is clear causes 1601 * iterate to re-test the 'current' element. If ATEDISK is 1602 * set, iterate will skip the 'current' element. 1603 */ 1604 error = 0; 1605 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 1606 if (cursor->flags & (HAMMER_CURSOR_ATEDISK | 1607 HAMMER_CURSOR_RETEST)) { 1608 error = hammer_btree_iterate(cursor); 1609 cursor->flags &= ~HAMMER_CURSOR_RETEST; 1610 if (error == 0) { 1611 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 1612 hammer_cache_node(&cursor->ip->cache[1], 1613 cursor->node); 1614 } else if (error == ENOENT) { 1615 cursor->flags |= HAMMER_CURSOR_DISKEOF | 1616 HAMMER_CURSOR_ATEDISK; 1617 error = 0; 1618 } 1619 } 1620 } 1621 1622 /* 1623 * If the generation changed the backend has deleted or committed 1624 * one or more memory records since our last check. 1625 * 1626 * When this case occurs if the disk cursor is > current memory record 1627 * or the disk cursor is at EOF, we must re-seek the disk-cursor. 1628 * Since the cursor is ahead it must have not yet been eaten (if 1629 * not at eof anyway). (XXX data offset case?) 1630 * 1631 * NOTE: we are not doing a full check here. That will be handled 1632 * later on. 1633 * 1634 * If we have exhausted all memory records we do not have to do any 1635 * further seeks. 1636 */ 1637 while (cursor->rec_generation != cursor->ip->rec_generation && 1638 error == 0) { 1639 hkprintf("Debug: generation changed during scan @ino=%016jx\n", 1640 (intmax_t)cursor->ip->obj_id); 1641 cursor->rec_generation = cursor->ip->rec_generation; 1642 if (cursor->flags & HAMMER_CURSOR_MEMEOF) 1643 break; 1644 if (cursor->flags & HAMMER_CURSOR_DISKEOF) { 1645 r = 1; 1646 } else { 1647 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0); 1648 elm = &cursor->node->ondisk->elms[cursor->index]; 1649 r = hammer_btree_cmp(&elm->base, 1650 &cursor->iprec->leaf.base); 1651 } 1652 1653 /* 1654 * Do we re-seek the media cursor? 1655 */ 1656 if (r > 0) { 1657 if (_hammer_ip_reseek(cursor)) 1658 goto again; 1659 } 1660 } 1661 1662 /* 1663 * We can now safely get the next in-memory record. We cannot 1664 * block here. 1665 * 1666 * hammer_rec_scan_cmp: Is the record still in our general range, 1667 * (non-inclusive of snapshot exclusions)? 1668 * hammer_rec_scan_callback: Is the record in our snapshot? 1669 */ 1670 tmprec = NULL; 1671 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) { 1672 /* 1673 * If the current memory record was eaten then get the next 1674 * one. Stale records are skipped. 1675 */ 1676 if (cursor->flags & HAMMER_CURSOR_ATEMEM) { 1677 tmprec = cursor->iprec; 1678 cursor->iprec = NULL; 1679 rec = hammer_rec_rb_tree_RB_NEXT(tmprec); 1680 while (rec) { 1681 if (hammer_rec_scan_cmp(rec, cursor) != 0) 1682 break; 1683 if (hammer_rec_scan_callback(rec, cursor) != 0) 1684 break; 1685 rec = hammer_rec_rb_tree_RB_NEXT(rec); 1686 } 1687 if (cursor->iprec) { 1688 KKASSERT(cursor->iprec == rec); 1689 cursor->flags &= ~HAMMER_CURSOR_ATEMEM; 1690 } else { 1691 cursor->flags |= HAMMER_CURSOR_MEMEOF; 1692 } 1693 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1694 } 1695 } 1696 1697 /* 1698 * MEMORY RECORD VALIDITY TEST 1699 * 1700 * (We still can't block, which is why tmprec is being held so 1701 * long). 1702 * 1703 * If the memory record is no longer valid we skip it. It may 1704 * have been deleted by the frontend. If it was deleted or 1705 * committed by the backend the generation change re-seeked the 1706 * disk cursor and the record will be present there. 1707 */ 1708 if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) { 1709 KKASSERT(cursor->iprec); 1710 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0); 1711 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) { 1712 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1713 if (tmprec) 1714 hammer_rel_mem_record(tmprec); 1715 goto again; 1716 } 1717 } 1718 if (tmprec) 1719 hammer_rel_mem_record(tmprec); 1720 1721 /* 1722 * Extract either the disk or memory record depending on their 1723 * relative position. 1724 */ 1725 error = 0; 1726 switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) { 1727 case 0: 1728 /* 1729 * Both entries valid. Compare the entries and nominally 1730 * return the first one in the sort order. Numerous cases 1731 * require special attention, however. 1732 */ 1733 elm = &cursor->node->ondisk->elms[cursor->index]; 1734 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base); 1735 1736 /* 1737 * If the two entries differ only by their key (-2/2) or 1738 * create_tid (-1/1), and are DATA records, we may have a 1739 * nominal match. We have to calculate the base file 1740 * offset of the data. 1741 */ 1742 if (r <= 2 && r >= -2 && r != 0 && 1743 cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE && 1744 cursor->iprec->type == HAMMER_MEM_RECORD_DATA) { 1745 int64_t base1 = elm->leaf.base.key - elm->leaf.data_len; 1746 int64_t base2 = cursor->iprec->leaf.base.key - 1747 cursor->iprec->leaf.data_len; 1748 if (base1 == base2) 1749 r = 0; 1750 } 1751 1752 if (r < 0) { 1753 error = hammer_btree_extract_leaf(cursor); 1754 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1755 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1756 break; 1757 } 1758 1759 /* 1760 * If the entries match exactly the memory entry is either 1761 * an on-disk directory entry deletion or a bulk data 1762 * overwrite. If it is a directory entry deletion we eat 1763 * both entries. 1764 * 1765 * For the bulk-data overwrite case it is possible to have 1766 * visibility into both, which simply means the syncer 1767 * hasn't gotten around to doing the delete+insert sequence 1768 * on the B-Tree. Use the memory entry and throw away the 1769 * on-disk entry. 1770 * 1771 * If the in-memory record is not either of these we 1772 * probably caught the syncer while it was syncing it to 1773 * the media. Since we hold a shared lock on the cursor, 1774 * the in-memory record had better be marked deleted at 1775 * this point. 1776 */ 1777 if (r == 0) { 1778 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) { 1779 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1780 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1781 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1782 goto again; 1783 } 1784 } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) { 1785 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1786 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1787 } 1788 /* fall through to memory entry */ 1789 } else { 1790 hpanic("duplicate mem/B-Tree entry %p %d %08x", 1791 cursor->iprec, 1792 cursor->iprec->type, 1793 cursor->iprec->flags); 1794 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1795 goto again; 1796 } 1797 } 1798 /* fall through to the memory entry */ 1799 case HAMMER_CURSOR_ATEDISK: 1800 /* 1801 * Only the memory entry is valid. 1802 */ 1803 cursor->leaf = &cursor->iprec->leaf; 1804 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1805 cursor->flags |= HAMMER_CURSOR_LASTWASMEM; 1806 1807 /* 1808 * If the memory entry is an on-disk deletion we should have 1809 * also had found a B-Tree record. If the backend beat us 1810 * to it it would have interlocked the cursor and we should 1811 * have seen the in-memory record marked DELETED_FE. 1812 */ 1813 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL && 1814 (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1815 hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x", 1816 cursor->iprec, 1817 cursor->iprec->flags); 1818 } 1819 break; 1820 case HAMMER_CURSOR_ATEMEM: 1821 /* 1822 * Only the disk entry is valid 1823 */ 1824 error = hammer_btree_extract_leaf(cursor); 1825 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1826 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1827 break; 1828 default: 1829 /* 1830 * Neither entry is valid 1831 * 1832 * XXX error not set properly 1833 */ 1834 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1835 cursor->leaf = NULL; 1836 error = ENOENT; 1837 break; 1838 } 1839 return(error); 1840 } 1841 1842 /* 1843 * Resolve the cursor->data pointer for the current cursor position in 1844 * a merged iteration. 1845 */ 1846 int 1847 hammer_ip_resolve_data(hammer_cursor_t cursor) 1848 { 1849 hammer_record_t record; 1850 int error; 1851 1852 if (hammer_cursor_inmem(cursor)) { 1853 /* 1854 * The data associated with an in-memory record is usually 1855 * kmalloced, but reserve-ahead data records will have an 1856 * on-disk reference. 1857 * 1858 * NOTE: Reserve-ahead data records must be handled in the 1859 * context of the related high level buffer cache buffer 1860 * to interlock against async writes. 1861 * 1862 * NOTE: We might catch a direct write in-progress, in which 1863 * case we must wait for it to complete. The wait 1864 * function will also clean out any buffer aliases. 1865 * 1866 * (In fact, it is possible that the write had not 1867 * even started yet). 1868 */ 1869 record = cursor->iprec; 1870 cursor->data = record->data; 1871 error = 0; 1872 if (cursor->data == NULL) { 1873 hammer_io_direct_wait(record); 1874 KKASSERT(record->leaf.base.rec_type == 1875 HAMMER_RECTYPE_DATA); 1876 cursor->data = hammer_bread_ext(cursor->trans->hmp, 1877 record->leaf.data_offset, 1878 record->leaf.data_len, 1879 &error, 1880 &cursor->data_buffer); 1881 } 1882 } else { 1883 /* 1884 * Loading leaf here isn't necessary if it's guaranteed that 1885 * the cursor is at a leaf node (which basically should be) 1886 * because hammer_btree_extract_data() does that. 1887 */ 1888 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf; 1889 error = hammer_btree_extract_data(cursor); 1890 } 1891 return(error); 1892 } 1893 1894 /* 1895 * Backend truncation / record replacement - delete records in range. 1896 * 1897 * Delete all records within the specified range for inode ip. In-memory 1898 * records still associated with the frontend are ignored. 1899 * 1900 * If truncating is non-zero in-memory records associated with the back-end 1901 * are ignored. If truncating is > 1 we can return EWOULDBLOCK. 1902 * 1903 * NOTES: 1904 * 1905 * * An unaligned range will cause new records to be added to cover 1906 * the edge cases. (XXX not implemented yet). 1907 * 1908 * * Replacement via reservations (see hammer_ip_sync_record_cursor()) 1909 * also do not deal with unaligned ranges. 1910 * 1911 * * ran_end is inclusive (e.g. 0,1023 instead of 0,1024). 1912 * 1913 * * Record keys for regular file data have to be special-cased since 1914 * they indicate the end of the range (key = base + bytes). 1915 * 1916 * * This function may be asked to delete ridiculously huge ranges, for 1917 * example if someone truncates or removes a 1TB regular file. We 1918 * must be very careful on restarts and we may have to stop w/ 1919 * EWOULDBLOCK to avoid blowing out the buffer cache. 1920 */ 1921 int 1922 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip, 1923 int64_t ran_beg, int64_t ran_end, int truncating) 1924 { 1925 hammer_transaction_t trans = cursor->trans; 1926 hammer_btree_leaf_elm_t leaf; 1927 int error; 1928 int64_t off; 1929 int64_t tmp64; 1930 1931 KKASSERT(trans->type == HAMMER_TRANS_FLS); 1932 retry: 1933 hammer_normalize_cursor(cursor); 1934 cursor->key_beg.localization = ip->obj_localization | 1935 HAMMER_LOCALIZE_MISC; 1936 cursor->key_beg.obj_id = ip->obj_id; 1937 cursor->key_beg.create_tid = 0; 1938 cursor->key_beg.delete_tid = 0; 1939 cursor->key_beg.obj_type = 0; 1940 1941 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1942 cursor->key_beg.key = ran_beg; 1943 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB; 1944 } else { 1945 /* 1946 * The key in the B-Tree is (base+bytes), so the first possible 1947 * matching key is ran_beg + 1. 1948 */ 1949 cursor->key_beg.key = ran_beg + 1; 1950 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA; 1951 } 1952 1953 cursor->key_end = cursor->key_beg; 1954 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1955 cursor->key_end.key = ran_end; 1956 } else { 1957 tmp64 = ran_end + MAXPHYS + 1; /* work around GCC-4 bug */ 1958 if (tmp64 < ran_end) 1959 cursor->key_end.key = HAMMER_MAX_KEY; 1960 else 1961 cursor->key_end.key = ran_end + MAXPHYS + 1; 1962 } 1963 1964 cursor->asof = ip->obj_asof; 1965 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 1966 cursor->flags |= HAMMER_CURSOR_ASOF; 1967 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 1968 cursor->flags |= HAMMER_CURSOR_BACKEND; 1969 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE; 1970 1971 error = hammer_ip_first(cursor); 1972 1973 /* 1974 * Iterate through matching records and mark them as deleted. 1975 */ 1976 while (error == 0) { 1977 leaf = cursor->leaf; 1978 1979 KKASSERT(leaf->base.delete_tid == 0); 1980 KKASSERT(leaf->base.obj_id == ip->obj_id); 1981 1982 /* 1983 * There may be overlap cases for regular file data. Also 1984 * remember the key for a regular file record is (base + len), 1985 * NOT (base). 1986 * 1987 * Note that due to duplicates (mem & media) allowed by 1988 * DELETE_VISIBILITY, off can wind up less then ran_beg. 1989 */ 1990 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) { 1991 off = leaf->base.key - leaf->data_len; 1992 /* 1993 * Check the left edge case. We currently do not 1994 * split existing records. 1995 */ 1996 if (off < ran_beg && leaf->base.key > ran_beg) { 1997 hpanic("hammer left edge case %016jx %d", 1998 (intmax_t)leaf->base.key, 1999 leaf->data_len); 2000 } 2001 2002 /* 2003 * Check the right edge case. Note that the 2004 * record can be completely out of bounds, which 2005 * terminates the search. 2006 * 2007 * base->key is exclusive of the right edge while 2008 * ran_end is inclusive of the right edge. The 2009 * (key - data_len) left boundary is inclusive. 2010 * 2011 * XXX theory-check this test at some point, are 2012 * we missing a + 1 somewhere? Note that ran_end 2013 * could overflow. 2014 */ 2015 if (leaf->base.key - 1 > ran_end) { 2016 if (leaf->base.key - leaf->data_len > ran_end) 2017 break; 2018 hpanic("hammer right edge case"); 2019 } 2020 } else { 2021 off = leaf->base.key; 2022 } 2023 2024 /* 2025 * Delete the record. When truncating we do not delete 2026 * in-memory (data) records because they represent data 2027 * written after the truncation. 2028 * 2029 * This will also physically destroy the B-Tree entry and 2030 * data if the retention policy dictates. The function 2031 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next() 2032 * to retest the new 'current' element. 2033 */ 2034 if (truncating == 0 || hammer_cursor_ondisk(cursor)) { 2035 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2036 /* 2037 * If we have built up too many meta-buffers we risk 2038 * deadlocking the kernel and must stop. This can 2039 * occur when deleting ridiculously huge files. 2040 * sync_trunc_off is updated so the next cycle does 2041 * not re-iterate records we have already deleted. 2042 * 2043 * This is only done with formal truncations. 2044 */ 2045 if (truncating > 1 && error == 0 && 2046 hammer_flusher_meta_limit(ip->hmp)) { 2047 ip->sync_trunc_off = off; 2048 error = EWOULDBLOCK; 2049 } 2050 } 2051 if (error) 2052 break; 2053 ran_beg = off; /* for restart */ 2054 error = hammer_ip_next(cursor); 2055 } 2056 if (cursor->node) 2057 hammer_cache_node(&ip->cache[1], cursor->node); 2058 2059 if (error == EDEADLK) { 2060 hammer_done_cursor(cursor); 2061 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2062 if (error == 0) 2063 goto retry; 2064 } 2065 if (error == ENOENT) 2066 error = 0; 2067 return(error); 2068 } 2069 2070 /* 2071 * This backend function deletes the specified record on-disk, similar to 2072 * delete_range but for a specific record. Unlike the exact deletions 2073 * used when deleting a directory entry this function uses an ASOF search 2074 * like delete_range. 2075 * 2076 * This function may be called with ip->obj_asof set for a slave snapshot, 2077 * so don't use it. We always delete non-historical records only. 2078 */ 2079 static int 2080 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip, 2081 hammer_btree_leaf_elm_t leaf) 2082 { 2083 hammer_transaction_t trans = cursor->trans; 2084 int error; 2085 2086 KKASSERT(trans->type == HAMMER_TRANS_FLS); 2087 retry: 2088 hammer_normalize_cursor(cursor); 2089 cursor->key_beg = leaf->base; 2090 cursor->asof = HAMMER_MAX_TID; 2091 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 2092 cursor->flags |= HAMMER_CURSOR_ASOF; 2093 cursor->flags |= HAMMER_CURSOR_BACKEND; 2094 cursor->flags &= ~HAMMER_CURSOR_INSERT; 2095 2096 error = hammer_btree_lookup(cursor); 2097 if (error == 0) { 2098 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2099 } 2100 if (error == EDEADLK) { 2101 hammer_done_cursor(cursor); 2102 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2103 if (error == 0) 2104 goto retry; 2105 } 2106 return(error); 2107 } 2108 2109 /* 2110 * This function deletes remaining auxillary records when an inode is 2111 * being deleted. This function explicitly does not delete the 2112 * inode record, directory entry, data, or db records. Those must be 2113 * properly disposed of prior to this call. 2114 */ 2115 int 2116 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp) 2117 { 2118 hammer_transaction_t trans = cursor->trans; 2119 hammer_btree_leaf_elm_t leaf __debugvar; 2120 int error; 2121 2122 KKASSERT(trans->type == HAMMER_TRANS_FLS); 2123 retry: 2124 hammer_normalize_cursor(cursor); 2125 cursor->key_beg.localization = ip->obj_localization | 2126 HAMMER_LOCALIZE_MISC; 2127 cursor->key_beg.obj_id = ip->obj_id; 2128 cursor->key_beg.create_tid = 0; 2129 cursor->key_beg.delete_tid = 0; 2130 cursor->key_beg.obj_type = 0; 2131 cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START; 2132 cursor->key_beg.key = HAMMER_MIN_KEY; 2133 2134 cursor->key_end = cursor->key_beg; 2135 cursor->key_end.rec_type = HAMMER_RECTYPE_MAX; 2136 cursor->key_end.key = HAMMER_MAX_KEY; 2137 2138 cursor->asof = ip->obj_asof; 2139 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 2140 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF; 2141 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 2142 cursor->flags |= HAMMER_CURSOR_BACKEND; 2143 2144 error = hammer_ip_first(cursor); 2145 2146 /* 2147 * Iterate through matching records and mark them as deleted. 2148 */ 2149 while (error == 0) { 2150 leaf = cursor->leaf; 2151 2152 KKASSERT(leaf->base.delete_tid == 0); 2153 2154 /* 2155 * Mark the record and B-Tree entry as deleted. This will 2156 * also physically delete the B-Tree entry, record, and 2157 * data if the retention policy dictates. The function 2158 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next() 2159 * to retest the new 'current' element. 2160 * 2161 * Directory entries (and delete-on-disk directory entries) 2162 * must be synced and cannot be deleted. 2163 */ 2164 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2165 ++*countp; 2166 if (error) 2167 break; 2168 error = hammer_ip_next(cursor); 2169 } 2170 if (cursor->node) 2171 hammer_cache_node(&ip->cache[1], cursor->node); 2172 if (error == EDEADLK) { 2173 hammer_done_cursor(cursor); 2174 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2175 if (error == 0) 2176 goto retry; 2177 } 2178 if (error == ENOENT) 2179 error = 0; 2180 return(error); 2181 } 2182 2183 /* 2184 * Delete the record at the current cursor. On success the cursor will 2185 * be positioned appropriately for an iteration but may no longer be at 2186 * a leaf node. 2187 * 2188 * This routine is only called from the backend. 2189 * 2190 * NOTE: This can return EDEADLK, requiring the caller to terminate the 2191 * cursor and retry. 2192 */ 2193 int 2194 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip, 2195 hammer_tid_t tid) 2196 { 2197 hammer_record_t iprec; 2198 int error; 2199 2200 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 2201 KKASSERT(tid != 0); 2202 2203 /* 2204 * In-memory (unsynchronized) records can simply be freed. This 2205 * only occurs in range iterations since all other records are 2206 * individually synchronized. Thus there should be no confusion with 2207 * the interlock. 2208 * 2209 * An in-memory record may be deleted before being committed to disk, 2210 * but could have been accessed in the mean time. The reservation 2211 * code will deal with the case. 2212 */ 2213 if (hammer_cursor_inmem(cursor)) { 2214 iprec = cursor->iprec; 2215 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0); 2216 iprec->flags |= HAMMER_RECF_DELETED_FE; 2217 iprec->flags |= HAMMER_RECF_DELETED_BE; 2218 KKASSERT(iprec->ip == ip); 2219 ++ip->rec_generation; 2220 return(0); 2221 } 2222 2223 /* 2224 * On-disk records are marked as deleted by updating their delete_tid. 2225 * This does not effect their position in the B-Tree (which is based 2226 * on their create_tid). 2227 * 2228 * Frontend B-Tree operations track inodes so we tell 2229 * hammer_delete_at_cursor() not to. 2230 */ 2231 error = hammer_btree_extract_leaf(cursor); 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_mount_t hmp; 2278 hammer_buffer_t data_buffer; 2279 hammer_off_t ndata_offset; 2280 hammer_tid_t high_tid; 2281 void *ndata; 2282 int error; 2283 int doprop; 2284 2285 trans = cursor->trans; 2286 hmp = trans->hmp; 2287 data_buffer = NULL; 2288 ndata_offset = 0; 2289 doprop = 0; 2290 2291 KKASSERT((cursor->flags & 2292 (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) == 2293 (HAMMER_CURSOR_BACKEND)); 2294 2295 hammer_sync_lock_sh(trans); 2296 2297 if (leaf->data_len) { 2298 ndata = hammer_alloc_data(trans, leaf->data_len, 2299 leaf->base.rec_type, 2300 &ndata_offset, &data_buffer, 2301 0, &error); 2302 if (ndata == NULL) { 2303 hammer_sync_unlock(trans); 2304 return (error); 2305 } 2306 leaf->data_offset = ndata_offset; 2307 hammer_modify_buffer_noundo(trans, data_buffer); 2308 2309 switch(mode) { 2310 case HAMMER_CREATE_MODE_UMIRROR: 2311 error = copyin(udata, ndata, leaf->data_len); 2312 if (error == 0) { 2313 if (hammer_crc_test_leaf(hmp->version, ndata, leaf) == 0) { 2314 hdkprintf("CRC DATA @ %016jx/%d MISMATCH ON PIPE\n", 2315 (intmax_t)ndata_offset, 2316 leaf->data_len); 2317 error = EINVAL; 2318 } else { 2319 error = hammer_cursor_localize_data( 2320 hmp, ndata, leaf); 2321 } 2322 } 2323 break; 2324 case HAMMER_CREATE_MODE_SYS: 2325 bcopy(udata, ndata, leaf->data_len); 2326 error = 0; 2327 hammer_crc_set_leaf(hmp->version, ndata, leaf); 2328 break; 2329 default: 2330 hpanic("bad mode %d", mode); 2331 break; /* NOT REACHED */ 2332 } 2333 hammer_modify_buffer_done(data_buffer); 2334 } else { 2335 leaf->data_offset = 0; 2336 error = 0; 2337 ndata = NULL; 2338 } 2339 if (error) 2340 goto failed; 2341 2342 /* 2343 * Do the insertion. This can fail with a EDEADLK or EALREADY 2344 */ 2345 cursor->flags |= HAMMER_CURSOR_INSERT; 2346 error = hammer_btree_lookup(cursor); 2347 if (error != ENOENT) { 2348 if (error == 0) 2349 error = EALREADY; 2350 goto failed; 2351 } 2352 error = hammer_btree_insert(cursor, leaf, &doprop); 2353 2354 /* 2355 * Cursor is left on current element, we want to skip it now. 2356 * (in case the caller is scanning) 2357 */ 2358 cursor->flags |= HAMMER_CURSOR_ATEDISK; 2359 cursor->flags &= ~HAMMER_CURSOR_INSERT; 2360 2361 /* 2362 * If the insertion happens to be creating (and not just replacing) 2363 * an inode we have to track it. 2364 */ 2365 if (error == 0 && 2366 leaf->base.rec_type == HAMMER_RECTYPE_INODE && 2367 leaf->base.delete_tid == 0) { 2368 hammer_modify_volume_field(trans, trans->rootvol, 2369 vol0_stat_inodes); 2370 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes; 2371 hammer_modify_volume_done(trans->rootvol); 2372 } 2373 2374 /* 2375 * vol0_next_tid must track the highest TID stored in the filesystem. 2376 * We do not need to generate undo for this update. 2377 */ 2378 high_tid = leaf->base.create_tid; 2379 if (high_tid < leaf->base.delete_tid) 2380 high_tid = leaf->base.delete_tid; 2381 if (trans->rootvol->ondisk->vol0_next_tid < high_tid) { 2382 hammer_modify_volume_noundo(trans, trans->rootvol); 2383 trans->rootvol->ondisk->vol0_next_tid = high_tid; 2384 hammer_modify_volume_done(trans->rootvol); 2385 } 2386 2387 /* 2388 * WARNING! cursor's leaf pointer may have changed after 2389 * do_propagation returns. 2390 */ 2391 if (error == 0 && doprop) 2392 hammer_btree_do_propagation(cursor, leaf); 2393 2394 failed: 2395 /* 2396 * Cleanup 2397 */ 2398 if (error && leaf->data_offset) { 2399 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len); 2400 2401 } 2402 hammer_sync_unlock(trans); 2403 if (data_buffer) 2404 hammer_rel_buffer(data_buffer, 0); 2405 return (error); 2406 } 2407 2408 /* 2409 * Delete the B-Tree element at the current cursor and do any necessary 2410 * mirror propagation. 2411 * 2412 * The cursor must be properly positioned for an iteration on return but 2413 * may be pointing at an internal element. 2414 * 2415 * An element can be un-deleted by passing a delete_tid of 0 with 2416 * HAMMER_DELETE_ADJUST. 2417 * 2418 * This function will store the number of bytes deleted in *stat_bytes 2419 * if stat_bytes is not NULL. 2420 */ 2421 int 2422 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags, 2423 hammer_tid_t delete_tid, uint32_t delete_ts, 2424 int track, int64_t *stat_bytes) 2425 { 2426 struct hammer_btree_leaf_elm save_leaf; 2427 hammer_transaction_t trans; 2428 hammer_btree_leaf_elm_t leaf; 2429 hammer_node_t node; 2430 hammer_btree_elm_t elm; 2431 hammer_off_t data_offset; 2432 int32_t data_len; 2433 int64_t bytes; 2434 int ndelete; 2435 int error; 2436 int icount; 2437 int doprop; 2438 2439 error = hammer_cursor_upgrade(cursor); 2440 if (error) 2441 return(error); 2442 2443 trans = cursor->trans; 2444 node = cursor->node; 2445 elm = &node->ondisk->elms[cursor->index]; 2446 leaf = &elm->leaf; 2447 KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD); 2448 2449 hammer_sync_lock_sh(trans); 2450 bytes = 0; 2451 doprop = 0; 2452 icount = 0; 2453 2454 /* 2455 * Adjust the delete_tid. Update the mirror_tid propagation field 2456 * as well. delete_tid can be 0 (undelete -- used by mirroring). 2457 */ 2458 if (delete_flags & HAMMER_DELETE_ADJUST) { 2459 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) { 2460 if (elm->leaf.base.delete_tid == 0 && delete_tid) 2461 icount = -1; 2462 if (elm->leaf.base.delete_tid && delete_tid == 0) 2463 icount = 1; 2464 } 2465 2466 hammer_modify_node(trans, node, elm, sizeof(*elm)); 2467 elm->leaf.base.delete_tid = delete_tid; 2468 elm->leaf.delete_ts = delete_ts; 2469 hammer_modify_node_done(node); 2470 2471 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) { 2472 hammer_modify_node_field(trans, node, mirror_tid); 2473 node->ondisk->mirror_tid = elm->leaf.base.delete_tid; 2474 hammer_modify_node_done(node); 2475 doprop = 1; 2476 if (hammer_debug_general & 0x0002) { 2477 hdkprintf("propagate %016jx @%016jx\n", 2478 (intmax_t)elm->leaf.base.delete_tid, 2479 (intmax_t)node->node_offset); 2480 } 2481 } 2482 2483 /* 2484 * Adjust for the iteration. We have deleted the current 2485 * element and want to clear ATEDISK so the iteration does 2486 * not skip the element after, which now becomes the current 2487 * element. This element must be re-tested if doing an 2488 * iteration, which is handled by the RETEST flag. 2489 */ 2490 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 2491 cursor->flags |= HAMMER_CURSOR_RETEST; 2492 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 2493 } 2494 2495 /* 2496 * An on-disk record cannot have the same delete_tid 2497 * as its create_tid. In a chain of record updates 2498 * this could result in a duplicate record. 2499 */ 2500 KKASSERT(elm->leaf.base.delete_tid != 2501 elm->leaf.base.create_tid); 2502 } 2503 2504 /* 2505 * Destroy the B-Tree element if asked (typically if a nohistory 2506 * file or mount, or when called by the pruning code). 2507 * 2508 * Adjust the ATEDISK flag to properly support iterations. 2509 */ 2510 if (delete_flags & HAMMER_DELETE_DESTROY) { 2511 data_offset = elm->leaf.data_offset; 2512 data_len = elm->leaf.data_len; 2513 if (doprop) { 2514 save_leaf = elm->leaf; 2515 leaf = &save_leaf; 2516 } 2517 if (elm->base.rec_type == HAMMER_RECTYPE_INODE && 2518 elm->leaf.base.delete_tid == 0) { 2519 icount = -1; 2520 } 2521 2522 error = hammer_btree_delete(cursor, &ndelete); 2523 if (error == 0) { 2524 /* 2525 * The deletion moves the next element (if any) to 2526 * the current element position. We must clear 2527 * ATEDISK so this element is not skipped and we 2528 * must set RETEST to force any iteration to re-test 2529 * the element. 2530 */ 2531 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 2532 cursor->flags |= HAMMER_CURSOR_RETEST; 2533 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 2534 } 2535 bytes += (ndelete * sizeof(struct hammer_node_ondisk)); 2536 2537 switch(HAMMER_ZONE(data_offset)) { 2538 case HAMMER_ZONE_LARGE_DATA: 2539 case HAMMER_ZONE_SMALL_DATA: 2540 case HAMMER_ZONE_META: 2541 hammer_blockmap_free(trans, 2542 data_offset, data_len); 2543 bytes += data_len; 2544 break; 2545 default: 2546 break; 2547 } 2548 } 2549 } 2550 2551 /* 2552 * Track inode count and next_tid. This is used by the mirroring 2553 * and PFS code. icount can be negative, zero, or positive. 2554 */ 2555 if (error == 0 && track) { 2556 if (icount) { 2557 hammer_modify_volume_field(trans, trans->rootvol, 2558 vol0_stat_inodes); 2559 trans->rootvol->ondisk->vol0_stat_inodes += icount; 2560 hammer_modify_volume_done(trans->rootvol); 2561 } 2562 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) { 2563 hammer_modify_volume_noundo(trans, trans->rootvol); 2564 trans->rootvol->ondisk->vol0_next_tid = delete_tid; 2565 hammer_modify_volume_done(trans->rootvol); 2566 } 2567 } 2568 2569 /* 2570 * mirror_tid propagation occurs if the node's mirror_tid had to be 2571 * updated while adjusting the delete_tid. 2572 * 2573 * This occurs when deleting even in nohistory mode, but does not 2574 * occur when pruning an already-deleted node. 2575 * 2576 * cursor->ip is NULL when called from the pruning, mirroring, 2577 * and pfs code. If non-NULL propagation will be conditionalized 2578 * on whether the PFS is in no-history mode or not. 2579 * 2580 * WARNING: cursor's leaf pointer may have changed after do_propagation 2581 * returns! 2582 */ 2583 if (doprop) { 2584 if (cursor->ip) 2585 hammer_btree_do_propagation(cursor, leaf); 2586 else 2587 hammer_btree_do_propagation(cursor, leaf); 2588 } 2589 if (stat_bytes) 2590 *stat_bytes = bytes; 2591 hammer_sync_unlock(trans); 2592 return (error); 2593 } 2594 2595 /* 2596 * Determine whether we can remove a directory. This routine checks whether 2597 * a directory is empty or not and enforces flush connectivity. 2598 * 2599 * Flush connectivity requires that we block if the target directory is 2600 * currently flushing, otherwise it may not end up in the same flush group. 2601 * 2602 * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure. 2603 */ 2604 int 2605 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip) 2606 { 2607 struct hammer_cursor cursor; 2608 int error; 2609 2610 /* 2611 * Check directory empty 2612 */ 2613 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip); 2614 2615 cursor.key_beg.localization = ip->obj_localization | 2616 hammer_dir_localization(ip); 2617 cursor.key_beg.obj_id = ip->obj_id; 2618 cursor.key_beg.create_tid = 0; 2619 cursor.key_beg.delete_tid = 0; 2620 cursor.key_beg.obj_type = 0; 2621 cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START; 2622 cursor.key_beg.key = HAMMER_MIN_KEY; 2623 2624 cursor.key_end = cursor.key_beg; 2625 cursor.key_end.rec_type = HAMMER_RECTYPE_MAX; 2626 cursor.key_end.key = HAMMER_MAX_KEY; 2627 2628 cursor.asof = ip->obj_asof; 2629 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF; 2630 2631 error = hammer_ip_first(&cursor); 2632 if (error == ENOENT) 2633 error = 0; 2634 else if (error == 0) 2635 error = ENOTEMPTY; 2636 hammer_done_cursor(&cursor); 2637 return(error); 2638 } 2639 2640 /* 2641 * Localize the data payload. Directory entries may need their 2642 * localization adjusted. 2643 */ 2644 static 2645 int 2646 hammer_cursor_localize_data(hammer_mount_t hmp, hammer_data_ondisk_t data, 2647 hammer_btree_leaf_elm_t leaf) 2648 { 2649 uint32_t localization; 2650 2651 if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) { 2652 localization = leaf->base.localization & 2653 HAMMER_LOCALIZE_PSEUDOFS_MASK; 2654 if (data->entry.localization != localization) { 2655 data->entry.localization = localization; 2656 hammer_crc_set_leaf(hmp->version, data, leaf); 2657 } 2658 } 2659 return(0); 2660 } 2661