1 /* 2 * Copyright (c) 2007-2008 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include "hammer.h" 36 37 static int hammer_mem_lookup(hammer_cursor_t cursor); 38 static int hammer_mem_first(hammer_cursor_t cursor); 39 static int hammer_frontend_trunc_callback(hammer_record_t record, 40 void *data __unused); 41 static int hammer_bulk_scan_callback(hammer_record_t record, void *data); 42 static int hammer_record_needs_overwrite_delete(hammer_record_t record); 43 static int hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip, 44 hammer_btree_leaf_elm_t leaf); 45 static int hammer_cursor_localize_data(hammer_data_ondisk_t data, 46 hammer_btree_leaf_elm_t leaf); 47 48 struct rec_trunc_info { 49 uint16_t rec_type; 50 int64_t trunc_off; 51 }; 52 53 struct hammer_bulk_info { 54 hammer_record_t record; 55 hammer_record_t conflict; 56 }; 57 58 /* 59 * Red-black tree support. Comparison code for insertion. 60 */ 61 static int 62 hammer_rec_rb_compare(hammer_record_t rec1, hammer_record_t rec2) 63 { 64 if (rec1->leaf.base.rec_type < rec2->leaf.base.rec_type) 65 return(-1); 66 if (rec1->leaf.base.rec_type > rec2->leaf.base.rec_type) 67 return(1); 68 69 if (rec1->leaf.base.key < rec2->leaf.base.key) 70 return(-1); 71 if (rec1->leaf.base.key > rec2->leaf.base.key) 72 return(1); 73 74 /* 75 * For search & insertion purposes records deleted by the 76 * frontend or deleted/committed by the backend are silently 77 * ignored. Otherwise pipelined insertions will get messed 78 * up. 79 * 80 * rec1 is greater then rec2 if rec1 is marked deleted. 81 * rec1 is less then rec2 if rec2 is marked deleted. 82 * 83 * Multiple deleted records may be present, do not return 0 84 * if both are marked deleted. 85 */ 86 if (rec1->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 87 HAMMER_RECF_COMMITTED)) { 88 return(1); 89 } 90 if (rec2->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 91 HAMMER_RECF_COMMITTED)) { 92 return(-1); 93 } 94 95 return(0); 96 } 97 98 /* 99 * Basic record comparison code similar to hammer_btree_cmp(). 100 * 101 * obj_id is not compared and may not yet be assigned in the record. 102 */ 103 static int 104 hammer_rec_cmp(hammer_base_elm_t elm, hammer_record_t rec) 105 { 106 if (elm->rec_type < rec->leaf.base.rec_type) 107 return(-3); 108 if (elm->rec_type > rec->leaf.base.rec_type) 109 return(3); 110 111 if (elm->key < rec->leaf.base.key) 112 return(-2); 113 if (elm->key > rec->leaf.base.key) 114 return(2); 115 116 /* 117 * Never match against an item deleted by the frontend 118 * or backend, or committed by the backend. 119 * 120 * elm is less then rec if rec is marked deleted. 121 */ 122 if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 123 HAMMER_RECF_COMMITTED)) { 124 return(-1); 125 } 126 return(0); 127 } 128 129 /* 130 * Ranged scan to locate overlapping record(s). This is used by 131 * hammer_ip_get_bulk() to locate an overlapping record. We have 132 * to use a ranged scan because the keys for data records with the 133 * same file base offset can be different due to differing data_len's. 134 * 135 * NOTE: The base file offset of a data record is (key - data_len), not (key). 136 */ 137 static int 138 hammer_rec_overlap_cmp(hammer_record_t rec, void *data) 139 { 140 struct hammer_bulk_info *info = data; 141 hammer_btree_leaf_elm_t leaf = &info->record->leaf; 142 143 if (rec->leaf.base.rec_type < leaf->base.rec_type) 144 return(-3); 145 if (rec->leaf.base.rec_type > leaf->base.rec_type) 146 return(3); 147 148 /* 149 * Overlap compare 150 */ 151 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) { 152 /* rec_beg >= leaf_end */ 153 if (rec->leaf.base.key - rec->leaf.data_len >= leaf->base.key) 154 return(2); 155 /* rec_end <= leaf_beg */ 156 if (rec->leaf.base.key <= leaf->base.key - leaf->data_len) 157 return(-2); 158 } else { 159 if (rec->leaf.base.key < leaf->base.key) 160 return(-2); 161 if (rec->leaf.base.key > leaf->base.key) 162 return(2); 163 } 164 165 /* 166 * We have to return 0 at this point, even if DELETED_FE is set, 167 * because returning anything else will cause the scan to ignore 168 * one of the branches when we really want it to check both. 169 */ 170 return(0); 171 } 172 173 /* 174 * RB_SCAN comparison code for hammer_mem_first(). The argument order 175 * is reversed so the comparison result has to be negated. key_beg and 176 * key_end are both range-inclusive. 177 * 178 * Localized deletions are not cached in-memory. 179 */ 180 static 181 int 182 hammer_rec_scan_cmp(hammer_record_t rec, void *data) 183 { 184 hammer_cursor_t cursor = data; 185 int r; 186 187 r = hammer_rec_cmp(&cursor->key_beg, rec); 188 if (r > 1) 189 return(-1); 190 r = hammer_rec_cmp(&cursor->key_end, rec); 191 if (r < -1) 192 return(1); 193 return(0); 194 } 195 196 /* 197 * This compare function is used when simply looking up key_beg. 198 */ 199 static 200 int 201 hammer_rec_find_cmp(hammer_record_t rec, void *data) 202 { 203 hammer_cursor_t cursor = data; 204 int r; 205 206 r = hammer_rec_cmp(&cursor->key_beg, rec); 207 if (r > 1) 208 return(-1); 209 if (r < -1) 210 return(1); 211 return(0); 212 } 213 214 /* 215 * Locate blocks within the truncation range. Partial blocks do not count. 216 */ 217 static 218 int 219 hammer_rec_trunc_cmp(hammer_record_t rec, void *data) 220 { 221 struct rec_trunc_info *info = data; 222 223 if (rec->leaf.base.rec_type < info->rec_type) 224 return(-1); 225 if (rec->leaf.base.rec_type > info->rec_type) 226 return(1); 227 228 switch(rec->leaf.base.rec_type) { 229 case HAMMER_RECTYPE_DB: 230 /* 231 * DB record key is not beyond the truncation point, retain. 232 */ 233 if (rec->leaf.base.key < info->trunc_off) 234 return(-1); 235 break; 236 case HAMMER_RECTYPE_DATA: 237 /* 238 * DATA record offset start is not beyond the truncation point, 239 * retain. 240 */ 241 if (rec->leaf.base.key - rec->leaf.data_len < info->trunc_off) 242 return(-1); 243 break; 244 default: 245 hpanic("unexpected record type"); 246 } 247 248 /* 249 * The record start is >= the truncation point, return match, 250 * the record should be destroyed. 251 */ 252 return(0); 253 } 254 255 RB_GENERATE(hammer_rec_rb_tree, hammer_record, rb_node, hammer_rec_rb_compare); 256 257 /* 258 * Allocate a record for the caller to finish filling in. The record is 259 * returned referenced. In order to manually set data call this function 260 * with data_len=0 and then manually set record->leaf.data_len and 261 * record->data later. 262 */ 263 hammer_record_t 264 hammer_alloc_mem_record(hammer_inode_t ip, int data_len) 265 { 266 hammer_record_t record; 267 hammer_mount_t hmp; 268 269 hmp = ip->hmp; 270 ++hammer_count_records; 271 record = kmalloc(sizeof(*record), hmp->m_misc, 272 M_WAITOK | M_ZERO | M_USE_RESERVE); 273 record->flush_state = HAMMER_FST_IDLE; 274 record->ip = ip; 275 record->leaf.base.btype = HAMMER_BTREE_TYPE_RECORD; 276 record->leaf.data_len = data_len; 277 hammer_ref(&record->lock); 278 279 if (data_len) { 280 record->data = kmalloc(data_len, hmp->m_misc, M_WAITOK | M_ZERO); 281 record->flags |= HAMMER_RECF_ALLOCDATA; 282 ++hammer_count_record_datas; 283 } 284 285 return (record); 286 } 287 288 void 289 hammer_wait_mem_record_ident(hammer_record_t record, const char *ident) 290 { 291 while (record->flush_state == HAMMER_FST_FLUSH) { 292 record->flags |= HAMMER_RECF_WANTED; 293 tsleep(record, 0, ident, 0); 294 } 295 } 296 297 /* 298 * Called from the backend, hammer_inode.c, after a record has been 299 * flushed to disk. The record has been exclusively locked by the 300 * caller and interlocked with BE. 301 * 302 * We clean up the state, unlock, and release the record (the record 303 * was referenced by the fact that it was in the HAMMER_FST_FLUSH state). 304 */ 305 void 306 hammer_flush_record_done(hammer_record_t record, int error) 307 { 308 hammer_inode_t target_ip; 309 310 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 311 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE); 312 313 /* 314 * If an error occured, the backend was unable to sync the 315 * record to its media. Leave the record intact. 316 */ 317 if (error) { 318 hammer_critical_error(record->ip->hmp, record->ip, error, 319 "while flushing record"); 320 } 321 322 --record->flush_group->refs; 323 record->flush_group = NULL; 324 325 /* 326 * Adjust the flush state and dependancy based on success or 327 * failure. 328 */ 329 if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) { 330 if ((target_ip = record->target_ip) != NULL) { 331 TAILQ_REMOVE(&target_ip->target_list, record, 332 target_entry); 333 record->target_ip = NULL; 334 hammer_test_inode(target_ip); 335 } 336 record->flush_state = HAMMER_FST_IDLE; 337 } else { 338 if (record->target_ip) { 339 record->flush_state = HAMMER_FST_SETUP; 340 hammer_test_inode(record->ip); 341 hammer_test_inode(record->target_ip); 342 } else { 343 record->flush_state = HAMMER_FST_IDLE; 344 } 345 } 346 record->flags &= ~HAMMER_RECF_INTERLOCK_BE; 347 348 /* 349 * Cleanup 350 */ 351 if (record->flags & HAMMER_RECF_WANTED) { 352 record->flags &= ~HAMMER_RECF_WANTED; 353 wakeup(record); 354 } 355 hammer_rel_mem_record(record); 356 } 357 358 /* 359 * Release a memory record. Records marked for deletion are immediately 360 * removed from the RB-Tree but otherwise left intact until the last ref 361 * goes away. 362 */ 363 void 364 hammer_rel_mem_record(hammer_record_t record) 365 { 366 hammer_mount_t hmp; 367 hammer_reserve_t resv; 368 hammer_inode_t ip; 369 hammer_inode_t target_ip; 370 int diddrop; 371 372 hammer_rel(&record->lock); 373 374 if (hammer_norefs(&record->lock)) { 375 /* 376 * Upon release of the last reference wakeup any waiters. 377 * The record structure may get destroyed so callers will 378 * loop up and do a relookup. 379 * 380 * WARNING! Record must be removed from RB-TREE before we 381 * might possibly block. hammer_test_inode() can block! 382 */ 383 ip = record->ip; 384 hmp = ip->hmp; 385 386 /* 387 * Upon release of the last reference a record marked deleted 388 * by the front or backend, or committed by the backend, 389 * is destroyed. 390 */ 391 if (record->flags & (HAMMER_RECF_DELETED_FE | 392 HAMMER_RECF_DELETED_BE | 393 HAMMER_RECF_COMMITTED)) { 394 KKASSERT(hammer_isactive(&ip->lock) > 0); 395 KKASSERT(record->flush_state != HAMMER_FST_FLUSH); 396 397 /* 398 * target_ip may have zero refs, we have to ref it 399 * to prevent it from being ripped out from under 400 * us. 401 */ 402 if ((target_ip = record->target_ip) != NULL) { 403 TAILQ_REMOVE(&target_ip->target_list, 404 record, target_entry); 405 record->target_ip = NULL; 406 hammer_ref(&target_ip->lock); 407 } 408 409 /* 410 * Remove the record from the RB-Tree 411 */ 412 if (record->flags & HAMMER_RECF_ONRBTREE) { 413 RB_REMOVE(hammer_rec_rb_tree, 414 &ip->rec_tree, 415 record); 416 record->flags &= ~HAMMER_RECF_ONRBTREE; 417 KKASSERT(ip->rsv_recs > 0); 418 if (RB_EMPTY(&ip->rec_tree)) { 419 ip->flags &= ~HAMMER_INODE_XDIRTY; 420 ip->sync_flags &= ~HAMMER_INODE_XDIRTY; 421 } 422 diddrop = 1; 423 } else { 424 diddrop = 0; 425 } 426 427 /* 428 * We must wait for any direct-IO to complete before 429 * we can destroy the record because the bio may 430 * have a reference to it. 431 */ 432 if (record->gflags & 433 (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) { 434 hammer_io_direct_wait(record); 435 } 436 437 /* 438 * Account for the completion after the direct IO 439 * has completed. 440 */ 441 if (diddrop) { 442 --hmp->rsv_recs; 443 --ip->rsv_recs; 444 hmp->rsv_databytes -= record->leaf.data_len; 445 446 if (RB_EMPTY(&ip->rec_tree)) 447 hammer_test_inode(ip); 448 if ((ip->flags & HAMMER_INODE_RECSW) && 449 ip->rsv_recs <= hammer_limit_inode_recs/2) { 450 ip->flags &= ~HAMMER_INODE_RECSW; 451 wakeup(&ip->rsv_recs); 452 } 453 } 454 455 /* 456 * Do this test after removing record from the RB-Tree. 457 */ 458 if (target_ip) { 459 hammer_test_inode(target_ip); 460 hammer_rel_inode(target_ip, 0); 461 } 462 463 if (record->flags & HAMMER_RECF_ALLOCDATA) { 464 --hammer_count_record_datas; 465 kfree(record->data, hmp->m_misc); 466 record->flags &= ~HAMMER_RECF_ALLOCDATA; 467 } 468 469 /* 470 * Release the reservation. 471 * 472 * If the record was not committed we can theoretically 473 * undo the reservation. However, doing so might 474 * create weird edge cases with the ordering of 475 * direct writes because the related buffer cache 476 * elements are per-vnode. So we don't try. 477 */ 478 if ((resv = record->resv) != NULL) { 479 /* XXX undo leaf.data_offset,leaf.data_len */ 480 hammer_blockmap_reserve_complete(hmp, resv); 481 record->resv = NULL; 482 } 483 record->data = NULL; 484 --hammer_count_records; 485 kfree(record, hmp->m_misc); 486 } 487 } 488 } 489 490 /* 491 * Record visibility depends on whether the record is being accessed by 492 * the backend or the frontend. Backend tests ignore the frontend delete 493 * flag. Frontend tests do NOT ignore the backend delete/commit flags and 494 * must also check for commit races. 495 * 496 * Return non-zero if the record is visible, zero if it isn't or if it is 497 * deleted. Returns 0 if the record has been comitted (unless the special 498 * delete-visibility flag is set). A committed record must be located 499 * via the media B-Tree. Returns non-zero if the record is good. 500 * 501 * If HAMMER_CURSOR_DELETE_VISIBILITY is set we allow deleted memory 502 * records to be returned. This is so pending deletions are detected 503 * when using an iterator to locate an unused hash key, or when we need 504 * to locate historical records on-disk to destroy. 505 */ 506 static __inline 507 int 508 hammer_ip_iterate_mem_good(hammer_cursor_t cursor, hammer_record_t record) 509 { 510 if (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) 511 return(1); 512 if (cursor->flags & HAMMER_CURSOR_BACKEND) { 513 if (record->flags & (HAMMER_RECF_DELETED_BE | 514 HAMMER_RECF_COMMITTED)) { 515 return(0); 516 } 517 } else { 518 if (record->flags & (HAMMER_RECF_DELETED_FE | 519 HAMMER_RECF_DELETED_BE | 520 HAMMER_RECF_COMMITTED)) { 521 return(0); 522 } 523 } 524 return(1); 525 } 526 527 /* 528 * This callback is used as part of the RB_SCAN function for in-memory 529 * records. We terminate it (return -1) as soon as we get a match. 530 * 531 * This routine is used by frontend code. 532 * 533 * The primary compare code does not account for ASOF lookups. This 534 * code handles that case as well as a few others. 535 */ 536 static 537 int 538 hammer_rec_scan_callback(hammer_record_t rec, void *data) 539 { 540 hammer_cursor_t cursor = data; 541 542 /* 543 * We terminate on success, so this should be NULL on entry. 544 */ 545 KKASSERT(cursor->iprec == NULL); 546 547 /* 548 * Skip if the record was marked deleted or committed. 549 */ 550 if (hammer_ip_iterate_mem_good(cursor, rec) == 0) 551 return(0); 552 553 /* 554 * Skip if not visible due to our as-of TID 555 */ 556 if (cursor->flags & HAMMER_CURSOR_ASOF) { 557 if (cursor->asof < rec->leaf.base.create_tid) 558 return(0); 559 if (rec->leaf.base.delete_tid && 560 cursor->asof >= rec->leaf.base.delete_tid) { 561 return(0); 562 } 563 } 564 565 /* 566 * ref the record. The record is protected from backend B-Tree 567 * interactions by virtue of the cursor's IP lock. 568 */ 569 hammer_ref(&rec->lock); 570 571 /* 572 * The record may have been deleted or committed while we 573 * were blocked. XXX remove? 574 */ 575 if (hammer_ip_iterate_mem_good(cursor, rec) == 0) { 576 hammer_rel_mem_record(rec); 577 return(0); 578 } 579 580 /* 581 * Set the matching record and stop the scan. 582 */ 583 cursor->iprec = rec; 584 return(-1); 585 } 586 587 588 /* 589 * Lookup an in-memory record given the key specified in the cursor. Works 590 * just like hammer_btree_lookup() but operates on an inode's in-memory 591 * record list. 592 * 593 * The lookup must fail if the record is marked for deferred deletion. 594 * 595 * The API for mem/btree_lookup() does not mess with the ATE/EOF bits. 596 */ 597 static 598 int 599 hammer_mem_lookup(hammer_cursor_t cursor) 600 { 601 KKASSERT(cursor->ip != NULL); 602 if (cursor->iprec) { 603 hammer_rel_mem_record(cursor->iprec); 604 cursor->iprec = NULL; 605 } 606 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_find_cmp, 607 hammer_rec_scan_callback, cursor); 608 609 return (cursor->iprec ? 0 : ENOENT); 610 } 611 612 /* 613 * hammer_mem_first() - locate the first in-memory record matching the 614 * cursor within the bounds of the key range. 615 * 616 * WARNING! API is slightly different from btree_first(). hammer_mem_first() 617 * will set ATEMEM the same as MEMEOF, and does not return any error. 618 */ 619 static 620 int 621 hammer_mem_first(hammer_cursor_t cursor) 622 { 623 KKASSERT(cursor->ip != NULL); 624 if (cursor->iprec) { 625 hammer_rel_mem_record(cursor->iprec); 626 cursor->iprec = NULL; 627 } 628 hammer_rec_rb_tree_RB_SCAN(&cursor->ip->rec_tree, hammer_rec_scan_cmp, 629 hammer_rec_scan_callback, cursor); 630 631 if (cursor->iprec) 632 cursor->flags &= ~(HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM); 633 else 634 cursor->flags |= HAMMER_CURSOR_MEMEOF | HAMMER_CURSOR_ATEMEM; 635 636 return (cursor->iprec ? 0 : ENOENT); 637 } 638 639 /************************************************************************ 640 * HAMMER IN-MEMORY RECORD FUNCTIONS * 641 ************************************************************************ 642 * 643 * These functions manipulate in-memory records. Such records typically 644 * exist prior to being committed to disk or indexed via the on-disk B-Tree. 645 */ 646 647 /* 648 * Add a directory entry (dip,ncp) which references inode (ip). 649 * 650 * Note that the low 32 bits of the namekey are set temporarily to create 651 * a unique in-memory record, and may be modified a second time when the 652 * record is synchronized to disk. In particular, the low 32 bits cannot be 653 * all 0's when synching to disk, which is not handled here. 654 * 655 * NOTE: bytes does not include any terminating \0 on name, and name might 656 * not be terminated. 657 */ 658 int 659 hammer_ip_add_direntry(hammer_transaction_t trans, 660 hammer_inode_t dip, const char *name, int bytes, 661 hammer_inode_t ip) 662 { 663 struct hammer_cursor cursor; 664 hammer_record_t record; 665 int error; 666 uint32_t max_iterations; 667 668 KKASSERT(dip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY); 669 670 record = hammer_alloc_mem_record(dip, HAMMER_ENTRY_SIZE(bytes)); 671 672 record->type = HAMMER_MEM_RECORD_ADD; 673 record->leaf.base.localization = dip->obj_localization | 674 hammer_dir_localization(dip); 675 record->leaf.base.obj_id = dip->obj_id; 676 record->leaf.base.key = hammer_direntry_namekey(dip, name, bytes, 677 &max_iterations); 678 record->leaf.base.rec_type = HAMMER_RECTYPE_DIRENTRY; 679 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 680 record->data->entry.obj_id = ip->obj_id; 681 record->data->entry.localization = ip->obj_localization; 682 bcopy(name, record->data->entry.name, bytes); 683 684 ++ip->ino_data.nlinks; 685 ip->ino_data.ctime = trans->time; 686 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY); 687 688 /* 689 * Find an unused namekey. Both the in-memory record tree and 690 * the B-Tree are checked. We do not want historically deleted 691 * names to create a collision as our iteration space may be limited, 692 * and since create_tid wouldn't match anyway an ASOF search 693 * must be used to locate collisions. 694 * 695 * delete-visibility is set so pending deletions do not give us 696 * a false-negative on our ability to use an iterator. 697 * 698 * The iterator must not rollover the key. Directory keys only 699 * use the positive key space. 700 */ 701 hammer_init_cursor(trans, &cursor, &dip->cache[1], dip); 702 cursor.key_beg = record->leaf.base; 703 cursor.flags |= HAMMER_CURSOR_ASOF; 704 cursor.flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 705 cursor.asof = ip->obj_asof; 706 707 while (hammer_ip_lookup(&cursor) == 0) { 708 ++record->leaf.base.key; 709 KKASSERT(record->leaf.base.key > 0); 710 cursor.key_beg.key = record->leaf.base.key; 711 if (--max_iterations == 0) { 712 hammer_rel_mem_record(record); 713 hmkprintf(trans->hmp, "Failed to find an unused namekey\n"); 714 error = ENOSPC; 715 goto failed; 716 } 717 } 718 719 /* 720 * The target inode and the directory entry are bound together. 721 */ 722 record->target_ip = ip; 723 record->flush_state = HAMMER_FST_SETUP; 724 TAILQ_INSERT_TAIL(&ip->target_list, record, target_entry); 725 726 /* 727 * The inode now has a dependancy and must be taken out of the idle 728 * state. An inode not in an idle state is given an extra reference. 729 * 730 * When transitioning to a SETUP state flag for an automatic reflush 731 * when the dependancies are disposed of if someone is waiting on 732 * the inode. 733 */ 734 if (ip->flush_state == HAMMER_FST_IDLE) { 735 hammer_ref(&ip->lock); 736 ip->flush_state = HAMMER_FST_SETUP; 737 if (ip->flags & HAMMER_INODE_FLUSHW) 738 ip->flags |= HAMMER_INODE_REFLUSH; 739 } 740 error = hammer_mem_add(record); 741 if (error == 0) { 742 dip->ino_data.mtime = trans->time; 743 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME); 744 } 745 failed: 746 hammer_done_cursor(&cursor); 747 return(error); 748 } 749 750 /* 751 * Delete the directory entry and update the inode link count. The 752 * cursor must be seeked to the directory entry record being deleted. 753 * 754 * The related inode should be share-locked by the caller. The caller is 755 * on the frontend. It could also be NULL indicating that the directory 756 * entry being removed has no related inode. 757 * 758 * This function can return EDEADLK requiring the caller to terminate 759 * the cursor, any locks, wait on the returned record, and retry. 760 */ 761 int 762 hammer_ip_del_direntry(hammer_transaction_t trans, 763 hammer_cursor_t cursor, hammer_inode_t dip, 764 hammer_inode_t ip) 765 { 766 hammer_record_t record; 767 int error; 768 769 if (hammer_cursor_inmem(cursor)) { 770 /* 771 * In-memory (unsynchronized) records can simply be freed. 772 * 773 * Even though the HAMMER_RECF_DELETED_FE flag is ignored 774 * by the backend, we must still avoid races against the 775 * backend potentially syncing the record to the media. 776 * 777 * We cannot call hammer_ip_delete_record(), that routine may 778 * only be called from the backend. 779 */ 780 record = cursor->iprec; 781 if (record->flags & (HAMMER_RECF_INTERLOCK_BE | 782 HAMMER_RECF_DELETED_BE | 783 HAMMER_RECF_COMMITTED)) { 784 KKASSERT(cursor->deadlk_rec == NULL); 785 hammer_ref(&record->lock); 786 cursor->deadlk_rec = record; 787 error = EDEADLK; 788 } else { 789 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD); 790 record->flags |= HAMMER_RECF_DELETED_FE; 791 error = 0; 792 } 793 } else { 794 /* 795 * If the record is on-disk we have to queue the deletion by 796 * the record's key. This also causes lookups to skip the 797 * record (lookups for the purposes of finding an unused 798 * directory key do not skip the record). 799 */ 800 KKASSERT(dip->flags & 801 (HAMMER_INODE_ONDISK | HAMMER_INODE_DONDISK)); 802 record = hammer_alloc_mem_record(dip, 0); 803 record->type = HAMMER_MEM_RECORD_DEL; 804 record->leaf.base = cursor->leaf->base; 805 KKASSERT(dip->obj_id == record->leaf.base.obj_id); 806 807 /* 808 * ip may be NULL, indicating the deletion of a directory 809 * entry which has no related inode. 810 */ 811 record->target_ip = ip; 812 if (ip) { 813 record->flush_state = HAMMER_FST_SETUP; 814 TAILQ_INSERT_TAIL(&ip->target_list, record, 815 target_entry); 816 } else { 817 record->flush_state = HAMMER_FST_IDLE; 818 } 819 820 /* 821 * The inode now has a dependancy and must be taken out of 822 * the idle state. An inode not in an idle state is given 823 * an extra reference. 824 * 825 * When transitioning to a SETUP state flag for an automatic 826 * reflush when the dependancies are disposed of if someone 827 * is waiting on the inode. 828 */ 829 if (ip && ip->flush_state == HAMMER_FST_IDLE) { 830 hammer_ref(&ip->lock); 831 ip->flush_state = HAMMER_FST_SETUP; 832 if (ip->flags & HAMMER_INODE_FLUSHW) 833 ip->flags |= HAMMER_INODE_REFLUSH; 834 } 835 836 error = hammer_mem_add(record); 837 } 838 839 /* 840 * One less link. The file may still be open in the OS even after 841 * all links have gone away. 842 * 843 * We have to terminate the cursor before syncing the inode to 844 * avoid deadlocking against ourselves. XXX this may no longer 845 * be true. 846 * 847 * If nlinks drops to zero and the vnode is inactive (or there is 848 * no vnode), call hammer_inode_unloadable_check() to zonk the 849 * inode. If we don't do this here the inode will not be destroyed 850 * on-media until we unmount. 851 */ 852 if (error == 0) { 853 if (ip) { 854 --ip->ino_data.nlinks; /* do before we might block */ 855 ip->ino_data.ctime = trans->time; 856 } 857 dip->ino_data.mtime = trans->time; 858 hammer_modify_inode(trans, dip, HAMMER_INODE_MTIME); 859 if (ip) { 860 hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY); 861 if (ip->ino_data.nlinks == 0 && 862 (ip->vp == NULL || (ip->vp->v_flag & VINACTIVE))) { 863 hammer_done_cursor(cursor); 864 hammer_inode_unloadable_check(ip, 1); 865 hammer_flush_inode(ip, 0); 866 } 867 } 868 869 } 870 return(error); 871 } 872 873 /* 874 * Add a record to an inode. 875 * 876 * The caller must allocate the record with hammer_alloc_mem_record(ip,len) and 877 * initialize the following additional fields that are not initialized by these 878 * functions. 879 * 880 * The related inode should be share-locked by the caller. The caller is 881 * on the frontend. 882 * 883 * record->leaf.base.key 884 * record->leaf.base.rec_type 885 * record->leaf.base.localization 886 */ 887 int 888 hammer_ip_add_record(hammer_transaction_t trans, hammer_record_t record) 889 { 890 hammer_inode_t ip = record->ip; 891 int error; 892 893 KKASSERT(record->leaf.base.localization != 0); 894 record->leaf.base.obj_id = ip->obj_id; 895 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 896 error = hammer_mem_add(record); 897 return(error); 898 } 899 900 /* 901 * Locate a pre-existing bulk record in memory. The caller wishes to 902 * replace the record with a new one. The existing record may have a 903 * different length (and thus a different key) so we have to use an 904 * overlap check function. 905 */ 906 static hammer_record_t 907 hammer_ip_get_bulk(hammer_record_t record) 908 { 909 struct hammer_bulk_info info; 910 hammer_inode_t ip = record->ip; 911 912 info.record = record; 913 info.conflict = NULL; 914 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_overlap_cmp, 915 hammer_bulk_scan_callback, &info); 916 917 return(info.conflict); /* may be NULL */ 918 } 919 920 /* 921 * Take records vetted by overlap_cmp. The first non-deleted record 922 * (if any) stops the scan. 923 */ 924 static int 925 hammer_bulk_scan_callback(hammer_record_t record, void *data) 926 { 927 struct hammer_bulk_info *info = data; 928 929 if (record->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE | 930 HAMMER_RECF_COMMITTED)) { 931 return(0); 932 } 933 hammer_ref(&record->lock); 934 info->conflict = record; 935 return(-1); /* stop scan */ 936 } 937 938 /* 939 * Reserve blockmap space placemarked with an in-memory record. 940 * 941 * This routine is called by the frontend in order to be able to directly 942 * flush a buffer cache buffer. The frontend has locked the related buffer 943 * cache buffers and we should be able to manipulate any overlapping 944 * in-memory records. 945 * 946 * The caller is responsible for adding the returned record and deleting 947 * the returned conflicting record (if any), typically by calling 948 * hammer_ip_replace_bulk() (via hammer_io_direct_write()). 949 */ 950 hammer_record_t 951 hammer_ip_add_bulk(hammer_inode_t ip, off_t file_offset, void *data, int bytes, 952 int *errorp) 953 { 954 hammer_record_t record; 955 hammer_dedup_cache_t dcp; 956 hammer_crc_t crc; 957 int zone; 958 959 /* 960 * Create a record to cover the direct write. The record cannot 961 * be added to the in-memory RB tree here as it might conflict 962 * with an existing memory record. See hammer_io_direct_write(). 963 * 964 * The backend is responsible for finalizing the space reserved in 965 * this record. 966 * 967 * XXX bytes not aligned, depend on the reservation code to 968 * align the reservation. 969 */ 970 record = hammer_alloc_mem_record(ip, 0); 971 zone = hammer_data_zone_index(bytes); 972 if (bytes == 0) 973 crc = 0; 974 else 975 crc = crc32(data, bytes); 976 977 if (hammer_live_dedup == 0) 978 goto nodedup; 979 if ((dcp = hammer_dedup_cache_lookup(ip->hmp, crc)) != NULL) { 980 struct hammer_dedup_cache tmp = *dcp; 981 982 record->resv = hammer_blockmap_reserve_dedup(ip->hmp, zone, 983 bytes, tmp.data_offset, errorp); 984 if (record->resv == NULL) 985 goto nodedup; 986 987 if (!hammer_dedup_validate(&tmp, zone, bytes, data)) { 988 hammer_blockmap_reserve_complete(ip->hmp, record->resv); 989 goto nodedup; 990 } 991 992 record->leaf.data_offset = tmp.data_offset; 993 record->flags |= HAMMER_RECF_DEDUPED; 994 } else { 995 nodedup: 996 record->resv = hammer_blockmap_reserve(ip->hmp, zone, bytes, 997 &record->leaf.data_offset, errorp); 998 if (record->resv == NULL) { 999 hdkprintf("reservation failed\n"); 1000 hammer_rel_mem_record(record); 1001 return(NULL); 1002 } 1003 } 1004 1005 record->type = HAMMER_MEM_RECORD_DATA; 1006 record->leaf.base.rec_type = HAMMER_RECTYPE_DATA; 1007 record->leaf.base.obj_type = ip->ino_leaf.base.obj_type; 1008 record->leaf.base.obj_id = ip->obj_id; 1009 record->leaf.base.key = file_offset + bytes; 1010 record->leaf.base.localization = ip->obj_localization | 1011 HAMMER_LOCALIZE_MISC; 1012 record->leaf.data_len = bytes; 1013 record->leaf.data_crc = crc; 1014 KKASSERT(*errorp == 0); 1015 1016 return(record); 1017 } 1018 1019 /* 1020 * Called by hammer_io_direct_write() prior to any possible completion 1021 * of the BIO to emplace the memory record associated with the I/O and 1022 * to replace any prior memory record which might still be active. 1023 * 1024 * Setting the FE deleted flag on the old record (if any) avoids any RB 1025 * tree insertion conflict, amoung other things. 1026 * 1027 * This has to be done prior to the caller completing any related buffer 1028 * cache I/O or a reinstantiation of the buffer may load data from the 1029 * old media location instead of the new media location. The holding 1030 * of the locked buffer cache buffer serves to interlock the record 1031 * replacement operation. 1032 */ 1033 void 1034 hammer_ip_replace_bulk(hammer_mount_t hmp, hammer_record_t record) 1035 { 1036 hammer_record_t conflict; 1037 int error __debugvar; 1038 1039 while ((conflict = hammer_ip_get_bulk(record)) != NULL) { 1040 if ((conflict->flags & HAMMER_RECF_INTERLOCK_BE) == 0) { 1041 conflict->flags |= HAMMER_RECF_DELETED_FE; 1042 break; 1043 } 1044 conflict->flags |= HAMMER_RECF_WANTED; 1045 tsleep(conflict, 0, "hmrrc3", 0); 1046 hammer_rel_mem_record(conflict); 1047 } 1048 error = hammer_mem_add(record); 1049 if (conflict) 1050 hammer_rel_mem_record(conflict); 1051 KKASSERT(error == 0); 1052 } 1053 1054 /* 1055 * Frontend truncation code. Scan in-memory records only. On-disk records 1056 * and records in a flushing state are handled by the backend. The vnops 1057 * setattr code will handle the block containing the truncation point. 1058 * 1059 * Partial blocks are not deleted. 1060 * 1061 * This code is only called on regular files. 1062 */ 1063 int 1064 hammer_ip_frontend_trunc(hammer_inode_t ip, off_t file_size) 1065 { 1066 struct rec_trunc_info info; 1067 1068 switch(ip->ino_data.obj_type) { 1069 case HAMMER_OBJTYPE_REGFILE: 1070 info.rec_type = HAMMER_RECTYPE_DATA; 1071 break; 1072 case HAMMER_OBJTYPE_DBFILE: 1073 info.rec_type = HAMMER_RECTYPE_DB; 1074 break; 1075 default: 1076 return(EINVAL); 1077 } 1078 info.trunc_off = file_size; 1079 hammer_rec_rb_tree_RB_SCAN(&ip->rec_tree, hammer_rec_trunc_cmp, 1080 hammer_frontend_trunc_callback, &info); 1081 return(0); 1082 } 1083 1084 /* 1085 * Scan callback for frontend records to destroy during a truncation. 1086 * We must ensure that DELETED_FE is set on the record or the frontend 1087 * will get confused in future read() calls. 1088 * 1089 * NOTE: DELETED_FE cannot be set while the record interlock (BE) is held. 1090 * In this rare case we must wait for the interlock to be cleared. 1091 * 1092 * NOTE: This function is only called on regular files. There are further 1093 * restrictions to the setting of DELETED_FE on directory records 1094 * undergoing a flush due to sensitive inode link count calculations. 1095 */ 1096 static int 1097 hammer_frontend_trunc_callback(hammer_record_t record, void *data __unused) 1098 { 1099 if (record->flags & HAMMER_RECF_DELETED_FE) 1100 return(0); 1101 #if 0 1102 if (record->flush_state == HAMMER_FST_FLUSH) 1103 return(0); 1104 #endif 1105 hammer_ref(&record->lock); 1106 while (record->flags & HAMMER_RECF_INTERLOCK_BE) 1107 hammer_wait_mem_record_ident(record, "hmmtrr"); 1108 record->flags |= HAMMER_RECF_DELETED_FE; 1109 hammer_rel_mem_record(record); 1110 return(0); 1111 } 1112 1113 /* 1114 * Return 1 if the caller must check for and delete existing records 1115 * before writing out a new data record. 1116 * 1117 * Return 0 if the caller can just insert the record into the B-Tree without 1118 * checking. 1119 */ 1120 static int 1121 hammer_record_needs_overwrite_delete(hammer_record_t record) 1122 { 1123 hammer_inode_t ip = record->ip; 1124 int64_t file_offset; 1125 int r; 1126 1127 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) 1128 file_offset = record->leaf.base.key; 1129 else 1130 file_offset = record->leaf.base.key - record->leaf.data_len; 1131 r = (file_offset < ip->save_trunc_off); 1132 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1133 if (ip->save_trunc_off <= record->leaf.base.key) 1134 ip->save_trunc_off = record->leaf.base.key + 1; 1135 } else { 1136 if (ip->save_trunc_off < record->leaf.base.key) 1137 ip->save_trunc_off = record->leaf.base.key; 1138 } 1139 return(r); 1140 } 1141 1142 /* 1143 * Backend code. Sync a record to the media. 1144 */ 1145 int 1146 hammer_ip_sync_record_cursor(hammer_cursor_t cursor, hammer_record_t record) 1147 { 1148 hammer_transaction_t trans = cursor->trans; 1149 int64_t file_offset; 1150 int bytes; 1151 void *bdata; 1152 int error; 1153 int doprop; 1154 1155 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 1156 KKASSERT(record->flags & HAMMER_RECF_INTERLOCK_BE); 1157 KKASSERT(record->leaf.base.localization != 0); 1158 1159 /* 1160 * Any direct-write related to the record must complete before we 1161 * can sync the record to the on-disk media. 1162 */ 1163 if (record->gflags & (HAMMER_RECG_DIRECT_IO | HAMMER_RECG_DIRECT_INVAL)) 1164 hammer_io_direct_wait(record); 1165 1166 /* 1167 * If this is a bulk-data record placemarker there may be an existing 1168 * record on-disk, indicating a data overwrite. If there is the 1169 * on-disk record must be deleted before we can insert our new record. 1170 * 1171 * We've synthesized this record and do not know what the create_tid 1172 * on-disk is, nor how much data it represents. 1173 * 1174 * Keep in mind that (key) for data records is (base_offset + len), 1175 * not (base_offset). Also, we only want to get rid of on-disk 1176 * records since we are trying to sync our in-memory record, call 1177 * hammer_ip_delete_range() with truncating set to 1 to make sure 1178 * it skips in-memory records. 1179 * 1180 * It is ok for the lookup to return ENOENT. 1181 * 1182 * NOTE OPTIMIZATION: sync_trunc_off is used to determine if we have 1183 * to call hammer_ip_delete_range() or not. This also means we must 1184 * update sync_trunc_off() as we write. 1185 */ 1186 if (record->type == HAMMER_MEM_RECORD_DATA && 1187 hammer_record_needs_overwrite_delete(record)) { 1188 file_offset = record->leaf.base.key - record->leaf.data_len; 1189 bytes = HAMMER_BUFSIZE_DOALIGN(record->leaf.data_len); 1190 KKASSERT((file_offset & HAMMER_BUFMASK) == 0); 1191 error = hammer_ip_delete_range( 1192 cursor, record->ip, 1193 file_offset, file_offset + bytes - 1, 1194 1); 1195 if (error && error != ENOENT) 1196 goto done; 1197 } 1198 1199 /* 1200 * If this is a general record there may be an on-disk version 1201 * that must be deleted before we can insert the new record. 1202 */ 1203 if (record->type == HAMMER_MEM_RECORD_GENERAL) { 1204 error = hammer_delete_general(cursor, record->ip, &record->leaf); 1205 if (error && error != ENOENT) 1206 goto done; 1207 } 1208 1209 /* 1210 * Setup the cursor. 1211 */ 1212 hammer_normalize_cursor(cursor); 1213 cursor->key_beg = record->leaf.base; 1214 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 1215 cursor->flags |= HAMMER_CURSOR_BACKEND; 1216 cursor->flags &= ~HAMMER_CURSOR_INSERT; 1217 1218 /* 1219 * Records can wind up on-media before the inode itself is on-media. 1220 * Flag the case. 1221 */ 1222 record->ip->flags |= HAMMER_INODE_DONDISK; 1223 1224 /* 1225 * If we are deleting a directory entry an exact match must be 1226 * found on-disk. 1227 */ 1228 if (record->type == HAMMER_MEM_RECORD_DEL) { 1229 error = hammer_btree_lookup(cursor); 1230 if (error == 0) { 1231 KKASSERT(cursor->iprec == NULL); 1232 error = hammer_ip_delete_record(cursor, record->ip, 1233 trans->tid); 1234 if (error == 0) { 1235 record->flags |= HAMMER_RECF_DELETED_BE | 1236 HAMMER_RECF_COMMITTED; 1237 ++record->ip->rec_generation; 1238 } 1239 } 1240 goto done; 1241 } 1242 1243 /* 1244 * We are inserting. 1245 * 1246 * Issue a lookup to position the cursor and locate the insertion 1247 * point. The target key should not exist. If we are creating a 1248 * directory entry we may have to iterate the low 32 bits of the 1249 * key to find an unused key. 1250 */ 1251 hammer_sync_lock_sh(trans); 1252 cursor->flags |= HAMMER_CURSOR_INSERT; 1253 error = hammer_btree_lookup(cursor); 1254 if (hammer_debug_inode) 1255 hdkprintf("DOINSERT LOOKUP %d\n", error); 1256 if (error == 0) { 1257 hdkprintf("duplicate rec at (%016jx)\n", 1258 (intmax_t)record->leaf.base.key); 1259 if (hammer_debug_critical) 1260 Debugger("duplicate record1"); 1261 error = EIO; 1262 } 1263 1264 if (error != ENOENT) 1265 goto done_unlock; 1266 1267 /* 1268 * Allocate the record and data. The result buffers will be 1269 * marked as being modified and further calls to 1270 * hammer_modify_buffer() will result in unneeded UNDO records. 1271 * 1272 * Support zero-fill records (data == NULL and data_len != 0) 1273 */ 1274 if (record->type == HAMMER_MEM_RECORD_DATA) { 1275 /* 1276 * The data portion of a bulk-data record has already been 1277 * committed to disk, we need only adjust the layer2 1278 * statistics in the same transaction as our B-Tree insert. 1279 */ 1280 KKASSERT(record->leaf.data_offset != 0); 1281 error = hammer_blockmap_finalize(trans, 1282 record->resv, 1283 record->leaf.data_offset, 1284 record->leaf.data_len); 1285 1286 if (hammer_live_dedup == 2 && 1287 (record->flags & HAMMER_RECF_DEDUPED) == 0) { 1288 hammer_dedup_cache_add(record->ip, &record->leaf); 1289 } 1290 } else if (record->data && record->leaf.data_len) { 1291 /* 1292 * Wholely cached record, with data. Allocate the data. 1293 */ 1294 bdata = hammer_alloc_data(trans, record->leaf.data_len, 1295 record->leaf.base.rec_type, 1296 &record->leaf.data_offset, 1297 &cursor->data_buffer, 1298 0, &error); 1299 if (bdata == NULL) 1300 goto done_unlock; 1301 hammer_crc_set_leaf(record->data, &record->leaf); 1302 hammer_modify_buffer_noundo(trans, cursor->data_buffer); 1303 bcopy(record->data, bdata, record->leaf.data_len); 1304 hammer_modify_buffer_done(cursor->data_buffer); 1305 } else { 1306 /* 1307 * Wholely cached record, without data. 1308 */ 1309 record->leaf.data_offset = 0; 1310 record->leaf.data_crc = 0; 1311 } 1312 1313 error = hammer_btree_insert(cursor, &record->leaf, &doprop); 1314 if (hammer_debug_inode && error) { 1315 hdkprintf("BTREE INSERT error %d @ %016jx:%d key %016jx\n", 1316 error, 1317 (intmax_t)cursor->node->node_offset, 1318 cursor->index, 1319 (intmax_t)record->leaf.base.key); 1320 } 1321 1322 /* 1323 * Our record is on-disk and we normally mark the in-memory version 1324 * as having been committed (and not BE-deleted). 1325 * 1326 * If the record represented a directory deletion but we had to 1327 * sync a valid directory entry to disk due to dependancies, 1328 * we must convert the record to a covering delete so the 1329 * frontend does not have visibility on the synced entry. 1330 * 1331 * WARNING: cursor's leaf pointer may have changed after do_propagation 1332 * returns! 1333 */ 1334 if (error == 0) { 1335 if (doprop) { 1336 hammer_btree_do_propagation(cursor, &record->leaf); 1337 } 1338 if (record->flags & HAMMER_RECF_CONVERT_DELETE) { 1339 /* 1340 * Must convert deleted directory entry add 1341 * to a directory entry delete. 1342 */ 1343 KKASSERT(record->type == HAMMER_MEM_RECORD_ADD); 1344 record->flags &= ~HAMMER_RECF_DELETED_FE; 1345 record->type = HAMMER_MEM_RECORD_DEL; 1346 KKASSERT(record->ip->obj_id == record->leaf.base.obj_id); 1347 KKASSERT(record->flush_state == HAMMER_FST_FLUSH); 1348 record->flags &= ~HAMMER_RECF_CONVERT_DELETE; 1349 KKASSERT((record->flags & (HAMMER_RECF_COMMITTED | 1350 HAMMER_RECF_DELETED_BE)) == 0); 1351 /* converted record is not yet committed */ 1352 /* hammer_flush_record_done takes care of the rest */ 1353 } else { 1354 /* 1355 * Everything went fine and we are now done with 1356 * this record. 1357 */ 1358 record->flags |= HAMMER_RECF_COMMITTED; 1359 ++record->ip->rec_generation; 1360 } 1361 } else { 1362 if (record->leaf.data_offset) { 1363 hammer_blockmap_free(trans, record->leaf.data_offset, 1364 record->leaf.data_len); 1365 } 1366 } 1367 done_unlock: 1368 hammer_sync_unlock(trans); 1369 done: 1370 return(error); 1371 } 1372 1373 /* 1374 * Add the record to the inode's rec_tree. The low 32 bits of a directory 1375 * entry's key is used to deal with hash collisions in the upper 32 bits. 1376 * A unique 64 bit key is generated in-memory and may be regenerated a 1377 * second time when the directory record is flushed to the on-disk B-Tree. 1378 * 1379 * A referenced record is passed to this function. This function 1380 * eats the reference. If an error occurs the record will be deleted. 1381 * 1382 * A copy of the temporary record->data pointer provided by the caller 1383 * will be made. 1384 */ 1385 int 1386 hammer_mem_add(hammer_record_t record) 1387 { 1388 hammer_mount_t hmp = record->ip->hmp; 1389 1390 /* 1391 * Make a private copy of record->data 1392 */ 1393 if (record->data) 1394 KKASSERT(record->flags & HAMMER_RECF_ALLOCDATA); 1395 1396 /* 1397 * Insert into the RB tree. A unique key should have already 1398 * been selected if this is a directory entry. 1399 */ 1400 if (RB_INSERT(hammer_rec_rb_tree, &record->ip->rec_tree, record)) { 1401 record->flags |= HAMMER_RECF_DELETED_FE; 1402 hammer_rel_mem_record(record); 1403 return (EEXIST); 1404 } 1405 ++hmp->rsv_recs; 1406 ++record->ip->rsv_recs; 1407 record->ip->hmp->rsv_databytes += record->leaf.data_len; 1408 record->flags |= HAMMER_RECF_ONRBTREE; 1409 hammer_modify_inode(NULL, record->ip, HAMMER_INODE_XDIRTY); 1410 hammer_rel_mem_record(record); 1411 return(0); 1412 } 1413 1414 /************************************************************************ 1415 * HAMMER INODE MERGED-RECORD FUNCTIONS * 1416 ************************************************************************ 1417 * 1418 * These functions augment the B-Tree scanning functions in hammer_btree.c 1419 * by merging in-memory records with on-disk records. 1420 */ 1421 1422 /* 1423 * Locate a particular record either in-memory or on-disk. 1424 * 1425 * NOTE: This is basically a standalone routine, hammer_ip_next() may 1426 * NOT be called to iterate results. 1427 */ 1428 int 1429 hammer_ip_lookup(hammer_cursor_t cursor) 1430 { 1431 int error; 1432 1433 /* 1434 * If the element is in-memory return it without searching the 1435 * on-disk B-Tree 1436 */ 1437 KKASSERT(cursor->ip); 1438 error = hammer_mem_lookup(cursor); 1439 if (error == 0) { 1440 cursor->leaf = &cursor->iprec->leaf; 1441 return(error); 1442 } 1443 if (error != ENOENT) 1444 return(error); 1445 1446 /* 1447 * If the inode has on-disk components search the on-disk B-Tree. 1448 */ 1449 if ((cursor->ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) == 0) 1450 return(error); 1451 error = hammer_btree_lookup(cursor); 1452 if (error == 0) 1453 error = hammer_btree_extract_leaf(cursor); 1454 return(error); 1455 } 1456 1457 /* 1458 * Helper for hammer_ip_first()/hammer_ip_next() 1459 * 1460 * NOTE: Both ATEDISK and DISKEOF will be set the same. This sets up 1461 * hammer_ip_first() for calling hammer_ip_next(), and sets up the re-seek 1462 * state if hammer_ip_next() needs to re-seek. 1463 */ 1464 static __inline 1465 int 1466 _hammer_ip_seek_btree(hammer_cursor_t cursor) 1467 { 1468 hammer_inode_t ip = cursor->ip; 1469 int error; 1470 1471 if (ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DONDISK)) { 1472 error = hammer_btree_lookup(cursor); 1473 if (error == ENOENT || error == EDEADLK) { 1474 if (hammer_debug_general & 0x2000) { 1475 hdkprintf("error %d node %p %016jx index %d\n", 1476 error, cursor->node, 1477 (intmax_t)cursor->node->node_offset, 1478 cursor->index); 1479 } 1480 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 1481 error = hammer_btree_iterate(cursor); 1482 } 1483 if (error == 0) { 1484 cursor->flags &= ~(HAMMER_CURSOR_DISKEOF | 1485 HAMMER_CURSOR_ATEDISK); 1486 } else { 1487 cursor->flags |= HAMMER_CURSOR_DISKEOF | 1488 HAMMER_CURSOR_ATEDISK; 1489 if (error == ENOENT) 1490 error = 0; 1491 } 1492 } else { 1493 cursor->flags |= HAMMER_CURSOR_DISKEOF | HAMMER_CURSOR_ATEDISK; 1494 error = 0; 1495 } 1496 return(error); 1497 } 1498 1499 /* 1500 * Helper for hammer_ip_next() 1501 * 1502 * The caller has determined that the media cursor is further along than the 1503 * memory cursor and must be reseeked after a generation number change. 1504 */ 1505 static 1506 int 1507 _hammer_ip_reseek(hammer_cursor_t cursor) 1508 { 1509 struct hammer_base_elm save; 1510 hammer_btree_elm_t elm; 1511 int error __debugvar; 1512 int r; 1513 int again = 0; 1514 1515 /* 1516 * Do the re-seek. 1517 */ 1518 hkprintf("Debug: re-seeked during scan @ino=%016jx\n", 1519 (intmax_t)cursor->ip->obj_id); 1520 save = cursor->key_beg; 1521 cursor->key_beg = cursor->iprec->leaf.base; 1522 error = _hammer_ip_seek_btree(cursor); 1523 KKASSERT(error == 0); 1524 cursor->key_beg = save; 1525 1526 /* 1527 * If the memory record was previous returned to 1528 * the caller and the media record matches 1529 * (-1/+1: only create_tid differs), then iterate 1530 * the media record to avoid a double result. 1531 */ 1532 if ((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0 && 1533 (cursor->flags & HAMMER_CURSOR_LASTWASMEM)) { 1534 elm = &cursor->node->ondisk->elms[cursor->index]; 1535 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base); 1536 if (cursor->flags & HAMMER_CURSOR_ASOF) { 1537 if (r >= -1 && r <= 1) { 1538 hkprintf("Debug: iterated after " 1539 "re-seek (asof r=%d)\n", r); 1540 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1541 again = 1; 1542 } 1543 } else { 1544 if (r == 0) { 1545 hkprintf("Debug: iterated after " 1546 "re-seek\n"); 1547 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1548 again = 1; 1549 } 1550 } 1551 } 1552 return(again); 1553 } 1554 1555 /* 1556 * Locate the first record within the cursor's key_beg/key_end range, 1557 * restricted to a particular inode. 0 is returned on success, ENOENT 1558 * if no records matched the requested range, or some other error. 1559 * 1560 * When 0 is returned hammer_ip_next() may be used to iterate additional 1561 * records within the requested range. 1562 * 1563 * This function can return EDEADLK, requiring the caller to terminate 1564 * the cursor and try again. 1565 */ 1566 1567 int 1568 hammer_ip_first(hammer_cursor_t cursor) 1569 { 1570 hammer_inode_t ip __debugvar = cursor->ip; 1571 int error; 1572 1573 KKASSERT(ip != NULL); 1574 1575 /* 1576 * Clean up fields and setup for merged scan 1577 */ 1578 cursor->flags &= ~HAMMER_CURSOR_RETEST; 1579 1580 /* 1581 * Search the in-memory record list (Red-Black tree). Unlike the 1582 * B-Tree search, mem_first checks for records in the range. 1583 * 1584 * This function will setup both ATEMEM and MEMEOF properly for 1585 * the ip iteration. ATEMEM will be set if MEMEOF is set. 1586 */ 1587 hammer_mem_first(cursor); 1588 1589 /* 1590 * Detect generation changes during blockages, including 1591 * blockages which occur on the initial btree search. 1592 */ 1593 cursor->rec_generation = cursor->ip->rec_generation; 1594 1595 /* 1596 * Initial search and result 1597 */ 1598 error = _hammer_ip_seek_btree(cursor); 1599 if (error == 0) 1600 error = hammer_ip_next(cursor); 1601 1602 return (error); 1603 } 1604 1605 /* 1606 * Retrieve the next record in a merged iteration within the bounds of the 1607 * cursor. This call may be made multiple times after the cursor has been 1608 * initially searched with hammer_ip_first(). 1609 * 1610 * There are numerous special cases in this code to deal with races between 1611 * in-memory records and on-media records. 1612 * 1613 * 0 is returned on success, ENOENT if no further records match the 1614 * requested range, or some other error code is returned. 1615 */ 1616 int 1617 hammer_ip_next(hammer_cursor_t cursor) 1618 { 1619 hammer_btree_elm_t elm; 1620 hammer_record_t rec; 1621 hammer_record_t tmprec; 1622 int error; 1623 int r; 1624 1625 again: 1626 /* 1627 * Get the next on-disk record 1628 * 1629 * NOTE: If we deleted the last on-disk record we had scanned 1630 * ATEDISK will be clear and RETEST will be set, forcing 1631 * a call to iterate. The fact that ATEDISK is clear causes 1632 * iterate to re-test the 'current' element. If ATEDISK is 1633 * set, iterate will skip the 'current' element. 1634 */ 1635 error = 0; 1636 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 1637 if (cursor->flags & (HAMMER_CURSOR_ATEDISK | 1638 HAMMER_CURSOR_RETEST)) { 1639 error = hammer_btree_iterate(cursor); 1640 cursor->flags &= ~HAMMER_CURSOR_RETEST; 1641 if (error == 0) { 1642 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 1643 hammer_cache_node(&cursor->ip->cache[1], 1644 cursor->node); 1645 } else if (error == ENOENT) { 1646 cursor->flags |= HAMMER_CURSOR_DISKEOF | 1647 HAMMER_CURSOR_ATEDISK; 1648 error = 0; 1649 } 1650 } 1651 } 1652 1653 /* 1654 * If the generation changed the backend has deleted or committed 1655 * one or more memory records since our last check. 1656 * 1657 * When this case occurs if the disk cursor is > current memory record 1658 * or the disk cursor is at EOF, we must re-seek the disk-cursor. 1659 * Since the cursor is ahead it must have not yet been eaten (if 1660 * not at eof anyway). (XXX data offset case?) 1661 * 1662 * NOTE: we are not doing a full check here. That will be handled 1663 * later on. 1664 * 1665 * If we have exhausted all memory records we do not have to do any 1666 * further seeks. 1667 */ 1668 while (cursor->rec_generation != cursor->ip->rec_generation && 1669 error == 0) { 1670 hkprintf("Debug: generation changed during scan @ino=%016jx\n", 1671 (intmax_t)cursor->ip->obj_id); 1672 cursor->rec_generation = cursor->ip->rec_generation; 1673 if (cursor->flags & HAMMER_CURSOR_MEMEOF) 1674 break; 1675 if (cursor->flags & HAMMER_CURSOR_DISKEOF) { 1676 r = 1; 1677 } else { 1678 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEDISK) == 0); 1679 elm = &cursor->node->ondisk->elms[cursor->index]; 1680 r = hammer_btree_cmp(&elm->base, 1681 &cursor->iprec->leaf.base); 1682 } 1683 1684 /* 1685 * Do we re-seek the media cursor? 1686 */ 1687 if (r > 0) { 1688 if (_hammer_ip_reseek(cursor)) 1689 goto again; 1690 } 1691 } 1692 1693 /* 1694 * We can now safely get the next in-memory record. We cannot 1695 * block here. 1696 * 1697 * hammer_rec_scan_cmp: Is the record still in our general range, 1698 * (non-inclusive of snapshot exclusions)? 1699 * hammer_rec_scan_callback: Is the record in our snapshot? 1700 */ 1701 tmprec = NULL; 1702 if ((cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) { 1703 /* 1704 * If the current memory record was eaten then get the next 1705 * one. Stale records are skipped. 1706 */ 1707 if (cursor->flags & HAMMER_CURSOR_ATEMEM) { 1708 tmprec = cursor->iprec; 1709 cursor->iprec = NULL; 1710 rec = hammer_rec_rb_tree_RB_NEXT(tmprec); 1711 while (rec) { 1712 if (hammer_rec_scan_cmp(rec, cursor) != 0) 1713 break; 1714 if (hammer_rec_scan_callback(rec, cursor) != 0) 1715 break; 1716 rec = hammer_rec_rb_tree_RB_NEXT(rec); 1717 } 1718 if (cursor->iprec) { 1719 KKASSERT(cursor->iprec == rec); 1720 cursor->flags &= ~HAMMER_CURSOR_ATEMEM; 1721 } else { 1722 cursor->flags |= HAMMER_CURSOR_MEMEOF; 1723 } 1724 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1725 } 1726 } 1727 1728 /* 1729 * MEMORY RECORD VALIDITY TEST 1730 * 1731 * (We still can't block, which is why tmprec is being held so 1732 * long). 1733 * 1734 * If the memory record is no longer valid we skip it. It may 1735 * have been deleted by the frontend. If it was deleted or 1736 * committed by the backend the generation change re-seeked the 1737 * disk cursor and the record will be present there. 1738 */ 1739 if (error == 0 && (cursor->flags & HAMMER_CURSOR_MEMEOF) == 0) { 1740 KKASSERT(cursor->iprec); 1741 KKASSERT((cursor->flags & HAMMER_CURSOR_ATEMEM) == 0); 1742 if (!hammer_ip_iterate_mem_good(cursor, cursor->iprec)) { 1743 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1744 if (tmprec) 1745 hammer_rel_mem_record(tmprec); 1746 goto again; 1747 } 1748 } 1749 if (tmprec) 1750 hammer_rel_mem_record(tmprec); 1751 1752 /* 1753 * Extract either the disk or memory record depending on their 1754 * relative position. 1755 */ 1756 error = 0; 1757 switch(cursor->flags & (HAMMER_CURSOR_ATEDISK | HAMMER_CURSOR_ATEMEM)) { 1758 case 0: 1759 /* 1760 * Both entries valid. Compare the entries and nominally 1761 * return the first one in the sort order. Numerous cases 1762 * require special attention, however. 1763 */ 1764 elm = &cursor->node->ondisk->elms[cursor->index]; 1765 r = hammer_btree_cmp(&elm->base, &cursor->iprec->leaf.base); 1766 1767 /* 1768 * If the two entries differ only by their key (-2/2) or 1769 * create_tid (-1/1), and are DATA records, we may have a 1770 * nominal match. We have to calculate the base file 1771 * offset of the data. 1772 */ 1773 if (r <= 2 && r >= -2 && r != 0 && 1774 cursor->ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE && 1775 cursor->iprec->type == HAMMER_MEM_RECORD_DATA) { 1776 int64_t base1 = elm->leaf.base.key - elm->leaf.data_len; 1777 int64_t base2 = cursor->iprec->leaf.base.key - 1778 cursor->iprec->leaf.data_len; 1779 if (base1 == base2) 1780 r = 0; 1781 } 1782 1783 if (r < 0) { 1784 error = hammer_btree_extract_leaf(cursor); 1785 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1786 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1787 break; 1788 } 1789 1790 /* 1791 * If the entries match exactly the memory entry is either 1792 * an on-disk directory entry deletion or a bulk data 1793 * overwrite. If it is a directory entry deletion we eat 1794 * both entries. 1795 * 1796 * For the bulk-data overwrite case it is possible to have 1797 * visibility into both, which simply means the syncer 1798 * hasn't gotten around to doing the delete+insert sequence 1799 * on the B-Tree. Use the memory entry and throw away the 1800 * on-disk entry. 1801 * 1802 * If the in-memory record is not either of these we 1803 * probably caught the syncer while it was syncing it to 1804 * the media. Since we hold a shared lock on the cursor, 1805 * the in-memory record had better be marked deleted at 1806 * this point. 1807 */ 1808 if (r == 0) { 1809 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL) { 1810 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1811 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1812 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1813 goto again; 1814 } 1815 } else if (cursor->iprec->type == HAMMER_MEM_RECORD_DATA) { 1816 if ((cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1817 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1818 } 1819 /* fall through to memory entry */ 1820 } else { 1821 hpanic("duplicate mem/B-Tree entry %p %d %08x", 1822 cursor->iprec, 1823 cursor->iprec->type, 1824 cursor->iprec->flags); 1825 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1826 goto again; 1827 } 1828 } 1829 /* fall through to the memory entry */ 1830 case HAMMER_CURSOR_ATEDISK: 1831 /* 1832 * Only the memory entry is valid. 1833 */ 1834 cursor->leaf = &cursor->iprec->leaf; 1835 cursor->flags |= HAMMER_CURSOR_ATEMEM; 1836 cursor->flags |= HAMMER_CURSOR_LASTWASMEM; 1837 1838 /* 1839 * If the memory entry is an on-disk deletion we should have 1840 * also had found a B-Tree record. If the backend beat us 1841 * to it it would have interlocked the cursor and we should 1842 * have seen the in-memory record marked DELETED_FE. 1843 */ 1844 if (cursor->iprec->type == HAMMER_MEM_RECORD_DEL && 1845 (cursor->flags & HAMMER_CURSOR_DELETE_VISIBILITY) == 0) { 1846 hpanic("del-on-disk with no B-Tree entry iprec %p flags %08x", 1847 cursor->iprec, 1848 cursor->iprec->flags); 1849 } 1850 break; 1851 case HAMMER_CURSOR_ATEMEM: 1852 /* 1853 * Only the disk entry is valid 1854 */ 1855 error = hammer_btree_extract_leaf(cursor); 1856 cursor->flags |= HAMMER_CURSOR_ATEDISK; 1857 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1858 break; 1859 default: 1860 /* 1861 * Neither entry is valid 1862 * 1863 * XXX error not set properly 1864 */ 1865 cursor->flags &= ~HAMMER_CURSOR_LASTWASMEM; 1866 cursor->leaf = NULL; 1867 error = ENOENT; 1868 break; 1869 } 1870 return(error); 1871 } 1872 1873 /* 1874 * Resolve the cursor->data pointer for the current cursor position in 1875 * a merged iteration. 1876 */ 1877 int 1878 hammer_ip_resolve_data(hammer_cursor_t cursor) 1879 { 1880 hammer_record_t record; 1881 int error; 1882 1883 if (hammer_cursor_inmem(cursor)) { 1884 /* 1885 * The data associated with an in-memory record is usually 1886 * kmalloced, but reserve-ahead data records will have an 1887 * on-disk reference. 1888 * 1889 * NOTE: Reserve-ahead data records must be handled in the 1890 * context of the related high level buffer cache buffer 1891 * to interlock against async writes. 1892 * 1893 * NOTE: We might catch a direct write in-progress, in which 1894 * case we must wait for it to complete. The wait 1895 * function will also clean out any buffer aliases. 1896 * 1897 * (In fact, it is possible that the write had not 1898 * even started yet). 1899 */ 1900 record = cursor->iprec; 1901 cursor->data = record->data; 1902 error = 0; 1903 if (cursor->data == NULL) { 1904 hammer_io_direct_wait(record); 1905 KKASSERT(record->leaf.base.rec_type == 1906 HAMMER_RECTYPE_DATA); 1907 cursor->data = hammer_bread_ext(cursor->trans->hmp, 1908 record->leaf.data_offset, 1909 record->leaf.data_len, 1910 &error, 1911 &cursor->data_buffer); 1912 } 1913 } else { 1914 /* 1915 * Loading leaf here isn't necessary if it's guaranteed that 1916 * the cursor is at a leaf node (which basically should be) 1917 * because hammer_btree_extract_data() does that. 1918 */ 1919 cursor->leaf = &cursor->node->ondisk->elms[cursor->index].leaf; 1920 error = hammer_btree_extract_data(cursor); 1921 } 1922 return(error); 1923 } 1924 1925 /* 1926 * Backend truncation / record replacement - delete records in range. 1927 * 1928 * Delete all records within the specified range for inode ip. In-memory 1929 * records still associated with the frontend are ignored. 1930 * 1931 * If truncating is non-zero in-memory records associated with the back-end 1932 * are ignored. If truncating is > 1 we can return EWOULDBLOCK. 1933 * 1934 * NOTES: 1935 * 1936 * * An unaligned range will cause new records to be added to cover 1937 * the edge cases. (XXX not implemented yet). 1938 * 1939 * * Replacement via reservations (see hammer_ip_sync_record_cursor()) 1940 * also do not deal with unaligned ranges. 1941 * 1942 * * ran_end is inclusive (e.g. 0,1023 instead of 0,1024). 1943 * 1944 * * Record keys for regular file data have to be special-cased since 1945 * they indicate the end of the range (key = base + bytes). 1946 * 1947 * * This function may be asked to delete ridiculously huge ranges, for 1948 * example if someone truncates or removes a 1TB regular file. We 1949 * must be very careful on restarts and we may have to stop w/ 1950 * EWOULDBLOCK to avoid blowing out the buffer cache. 1951 */ 1952 int 1953 hammer_ip_delete_range(hammer_cursor_t cursor, hammer_inode_t ip, 1954 int64_t ran_beg, int64_t ran_end, int truncating) 1955 { 1956 hammer_transaction_t trans = cursor->trans; 1957 hammer_btree_leaf_elm_t leaf; 1958 int error; 1959 int64_t off; 1960 int64_t tmp64; 1961 1962 KKASSERT(trans->type == HAMMER_TRANS_FLS); 1963 retry: 1964 hammer_normalize_cursor(cursor); 1965 cursor->key_beg.localization = ip->obj_localization | 1966 HAMMER_LOCALIZE_MISC; 1967 cursor->key_beg.obj_id = ip->obj_id; 1968 cursor->key_beg.create_tid = 0; 1969 cursor->key_beg.delete_tid = 0; 1970 cursor->key_beg.obj_type = 0; 1971 1972 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1973 cursor->key_beg.key = ran_beg; 1974 cursor->key_beg.rec_type = HAMMER_RECTYPE_DB; 1975 } else { 1976 /* 1977 * The key in the B-Tree is (base+bytes), so the first possible 1978 * matching key is ran_beg + 1. 1979 */ 1980 cursor->key_beg.key = ran_beg + 1; 1981 cursor->key_beg.rec_type = HAMMER_RECTYPE_DATA; 1982 } 1983 1984 cursor->key_end = cursor->key_beg; 1985 if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DBFILE) { 1986 cursor->key_end.key = ran_end; 1987 } else { 1988 tmp64 = ran_end + MAXPHYS + 1; /* work around GCC-4 bug */ 1989 if (tmp64 < ran_end) 1990 cursor->key_end.key = HAMMER_MAX_KEY; 1991 else 1992 cursor->key_end.key = ran_end + MAXPHYS + 1; 1993 } 1994 1995 cursor->asof = ip->obj_asof; 1996 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 1997 cursor->flags |= HAMMER_CURSOR_ASOF; 1998 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 1999 cursor->flags |= HAMMER_CURSOR_BACKEND; 2000 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE; 2001 2002 error = hammer_ip_first(cursor); 2003 2004 /* 2005 * Iterate through matching records and mark them as deleted. 2006 */ 2007 while (error == 0) { 2008 leaf = cursor->leaf; 2009 2010 KKASSERT(leaf->base.delete_tid == 0); 2011 KKASSERT(leaf->base.obj_id == ip->obj_id); 2012 2013 /* 2014 * There may be overlap cases for regular file data. Also 2015 * remember the key for a regular file record is (base + len), 2016 * NOT (base). 2017 * 2018 * Note that due to duplicates (mem & media) allowed by 2019 * DELETE_VISIBILITY, off can wind up less then ran_beg. 2020 */ 2021 if (leaf->base.rec_type == HAMMER_RECTYPE_DATA) { 2022 off = leaf->base.key - leaf->data_len; 2023 /* 2024 * Check the left edge case. We currently do not 2025 * split existing records. 2026 */ 2027 if (off < ran_beg && leaf->base.key > ran_beg) { 2028 hpanic("hammer left edge case %016jx %d", 2029 (intmax_t)leaf->base.key, 2030 leaf->data_len); 2031 } 2032 2033 /* 2034 * Check the right edge case. Note that the 2035 * record can be completely out of bounds, which 2036 * terminates the search. 2037 * 2038 * base->key is exclusive of the right edge while 2039 * ran_end is inclusive of the right edge. The 2040 * (key - data_len) left boundary is inclusive. 2041 * 2042 * XXX theory-check this test at some point, are 2043 * we missing a + 1 somewhere? Note that ran_end 2044 * could overflow. 2045 */ 2046 if (leaf->base.key - 1 > ran_end) { 2047 if (leaf->base.key - leaf->data_len > ran_end) 2048 break; 2049 hpanic("hammer right edge case"); 2050 } 2051 } else { 2052 off = leaf->base.key; 2053 } 2054 2055 /* 2056 * Delete the record. When truncating we do not delete 2057 * in-memory (data) records because they represent data 2058 * written after the truncation. 2059 * 2060 * This will also physically destroy the B-Tree entry and 2061 * data if the retention policy dictates. The function 2062 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next() 2063 * to retest the new 'current' element. 2064 */ 2065 if (truncating == 0 || hammer_cursor_ondisk(cursor)) { 2066 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2067 /* 2068 * If we have built up too many meta-buffers we risk 2069 * deadlocking the kernel and must stop. This can 2070 * occur when deleting ridiculously huge files. 2071 * sync_trunc_off is updated so the next cycle does 2072 * not re-iterate records we have already deleted. 2073 * 2074 * This is only done with formal truncations. 2075 */ 2076 if (truncating > 1 && error == 0 && 2077 hammer_flusher_meta_limit(ip->hmp)) { 2078 ip->sync_trunc_off = off; 2079 error = EWOULDBLOCK; 2080 } 2081 } 2082 if (error) 2083 break; 2084 ran_beg = off; /* for restart */ 2085 error = hammer_ip_next(cursor); 2086 } 2087 if (cursor->node) 2088 hammer_cache_node(&ip->cache[1], cursor->node); 2089 2090 if (error == EDEADLK) { 2091 hammer_done_cursor(cursor); 2092 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2093 if (error == 0) 2094 goto retry; 2095 } 2096 if (error == ENOENT) 2097 error = 0; 2098 return(error); 2099 } 2100 2101 /* 2102 * This backend function deletes the specified record on-disk, similar to 2103 * delete_range but for a specific record. Unlike the exact deletions 2104 * used when deleting a directory entry this function uses an ASOF search 2105 * like delete_range. 2106 * 2107 * This function may be called with ip->obj_asof set for a slave snapshot, 2108 * so don't use it. We always delete non-historical records only. 2109 */ 2110 static int 2111 hammer_delete_general(hammer_cursor_t cursor, hammer_inode_t ip, 2112 hammer_btree_leaf_elm_t leaf) 2113 { 2114 hammer_transaction_t trans = cursor->trans; 2115 int error; 2116 2117 KKASSERT(trans->type == HAMMER_TRANS_FLS); 2118 retry: 2119 hammer_normalize_cursor(cursor); 2120 cursor->key_beg = leaf->base; 2121 cursor->asof = HAMMER_MAX_TID; 2122 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 2123 cursor->flags |= HAMMER_CURSOR_ASOF; 2124 cursor->flags |= HAMMER_CURSOR_BACKEND; 2125 cursor->flags &= ~HAMMER_CURSOR_INSERT; 2126 2127 error = hammer_btree_lookup(cursor); 2128 if (error == 0) { 2129 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2130 } 2131 if (error == EDEADLK) { 2132 hammer_done_cursor(cursor); 2133 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2134 if (error == 0) 2135 goto retry; 2136 } 2137 return(error); 2138 } 2139 2140 /* 2141 * This function deletes remaining auxillary records when an inode is 2142 * being deleted. This function explicitly does not delete the 2143 * inode record, directory entry, data, or db records. Those must be 2144 * properly disposed of prior to this call. 2145 */ 2146 int 2147 hammer_ip_delete_clean(hammer_cursor_t cursor, hammer_inode_t ip, int *countp) 2148 { 2149 hammer_transaction_t trans = cursor->trans; 2150 hammer_btree_leaf_elm_t leaf __debugvar; 2151 int error; 2152 2153 KKASSERT(trans->type == HAMMER_TRANS_FLS); 2154 retry: 2155 hammer_normalize_cursor(cursor); 2156 cursor->key_beg.localization = ip->obj_localization | 2157 HAMMER_LOCALIZE_MISC; 2158 cursor->key_beg.obj_id = ip->obj_id; 2159 cursor->key_beg.create_tid = 0; 2160 cursor->key_beg.delete_tid = 0; 2161 cursor->key_beg.obj_type = 0; 2162 cursor->key_beg.rec_type = HAMMER_RECTYPE_CLEAN_START; 2163 cursor->key_beg.key = HAMMER_MIN_KEY; 2164 2165 cursor->key_end = cursor->key_beg; 2166 cursor->key_end.rec_type = HAMMER_RECTYPE_MAX; 2167 cursor->key_end.key = HAMMER_MAX_KEY; 2168 2169 cursor->asof = ip->obj_asof; 2170 cursor->flags &= ~HAMMER_CURSOR_INITMASK; 2171 cursor->flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF; 2172 cursor->flags |= HAMMER_CURSOR_DELETE_VISIBILITY; 2173 cursor->flags |= HAMMER_CURSOR_BACKEND; 2174 2175 error = hammer_ip_first(cursor); 2176 2177 /* 2178 * Iterate through matching records and mark them as deleted. 2179 */ 2180 while (error == 0) { 2181 leaf = cursor->leaf; 2182 2183 KKASSERT(leaf->base.delete_tid == 0); 2184 2185 /* 2186 * Mark the record and B-Tree entry as deleted. This will 2187 * also physically delete the B-Tree entry, record, and 2188 * data if the retention policy dictates. The function 2189 * will set HAMMER_CURSOR_RETEST to cause hammer_ip_next() 2190 * to retest the new 'current' element. 2191 * 2192 * Directory entries (and delete-on-disk directory entries) 2193 * must be synced and cannot be deleted. 2194 */ 2195 error = hammer_ip_delete_record(cursor, ip, trans->tid); 2196 ++*countp; 2197 if (error) 2198 break; 2199 error = hammer_ip_next(cursor); 2200 } 2201 if (cursor->node) 2202 hammer_cache_node(&ip->cache[1], cursor->node); 2203 if (error == EDEADLK) { 2204 hammer_done_cursor(cursor); 2205 error = hammer_init_cursor(trans, cursor, &ip->cache[1], ip); 2206 if (error == 0) 2207 goto retry; 2208 } 2209 if (error == ENOENT) 2210 error = 0; 2211 return(error); 2212 } 2213 2214 /* 2215 * Delete the record at the current cursor. On success the cursor will 2216 * be positioned appropriately for an iteration but may no longer be at 2217 * a leaf node. 2218 * 2219 * This routine is only called from the backend. 2220 * 2221 * NOTE: This can return EDEADLK, requiring the caller to terminate the 2222 * cursor and retry. 2223 */ 2224 int 2225 hammer_ip_delete_record(hammer_cursor_t cursor, hammer_inode_t ip, 2226 hammer_tid_t tid) 2227 { 2228 hammer_record_t iprec; 2229 int error; 2230 2231 KKASSERT(cursor->flags & HAMMER_CURSOR_BACKEND); 2232 KKASSERT(tid != 0); 2233 2234 /* 2235 * In-memory (unsynchronized) records can simply be freed. This 2236 * only occurs in range iterations since all other records are 2237 * individually synchronized. Thus there should be no confusion with 2238 * the interlock. 2239 * 2240 * An in-memory record may be deleted before being committed to disk, 2241 * but could have been accessed in the mean time. The reservation 2242 * code will deal with the case. 2243 */ 2244 if (hammer_cursor_inmem(cursor)) { 2245 iprec = cursor->iprec; 2246 KKASSERT((iprec->flags & HAMMER_RECF_INTERLOCK_BE) ==0); 2247 iprec->flags |= HAMMER_RECF_DELETED_FE; 2248 iprec->flags |= HAMMER_RECF_DELETED_BE; 2249 KKASSERT(iprec->ip == ip); 2250 ++ip->rec_generation; 2251 return(0); 2252 } 2253 2254 /* 2255 * On-disk records are marked as deleted by updating their delete_tid. 2256 * This does not effect their position in the B-Tree (which is based 2257 * on their create_tid). 2258 * 2259 * Frontend B-Tree operations track inodes so we tell 2260 * hammer_delete_at_cursor() not to. 2261 */ 2262 error = hammer_btree_extract_leaf(cursor); 2263 if (error == 0) { 2264 error = hammer_delete_at_cursor( 2265 cursor, 2266 HAMMER_DELETE_ADJUST | hammer_nohistory(ip), 2267 cursor->trans->tid, 2268 cursor->trans->time32, 2269 0, NULL); 2270 } 2271 return(error); 2272 } 2273 2274 /* 2275 * Used to write a generic record w/optional data to the media b-tree 2276 * when no inode context is available. Used by the mirroring and 2277 * snapshot code. 2278 * 2279 * Caller must set cursor->key_beg to leaf->base. The cursor must be 2280 * flagged for backend operation and not flagged ASOF (since we are 2281 * doing an insertion). 2282 * 2283 * This function will acquire the appropriate sync lock and will set 2284 * the cursor insertion flag for the operation, do the btree lookup, 2285 * and the insertion, and clear the insertion flag and sync lock before 2286 * returning. The cursor state will be such that the caller can continue 2287 * scanning (used by the mirroring code). 2288 * 2289 * mode: HAMMER_CREATE_MODE_UMIRROR copyin data, check crc 2290 * HAMMER_CREATE_MODE_SYS bcopy data, generate crc 2291 * 2292 * NOTE: EDEADLK can be returned. The caller must do deadlock handling and 2293 * retry. 2294 * 2295 * EALREADY can be returned if the record already exists (WARNING, 2296 * because ASOF cannot be used no check is made for illegal 2297 * duplicates). 2298 * 2299 * NOTE: Do not use the function for normal inode-related records as this 2300 * functions goes directly to the media and is not integrated with 2301 * in-memory records. 2302 */ 2303 int 2304 hammer_create_at_cursor(hammer_cursor_t cursor, hammer_btree_leaf_elm_t leaf, 2305 void *udata, int mode) 2306 { 2307 hammer_transaction_t trans; 2308 hammer_buffer_t data_buffer; 2309 hammer_off_t ndata_offset; 2310 hammer_tid_t high_tid; 2311 void *ndata; 2312 int error; 2313 int doprop; 2314 2315 trans = cursor->trans; 2316 data_buffer = NULL; 2317 ndata_offset = 0; 2318 doprop = 0; 2319 2320 KKASSERT((cursor->flags & 2321 (HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF)) == 2322 (HAMMER_CURSOR_BACKEND)); 2323 2324 hammer_sync_lock_sh(trans); 2325 2326 if (leaf->data_len) { 2327 ndata = hammer_alloc_data(trans, leaf->data_len, 2328 leaf->base.rec_type, 2329 &ndata_offset, &data_buffer, 2330 0, &error); 2331 if (ndata == NULL) { 2332 hammer_sync_unlock(trans); 2333 return (error); 2334 } 2335 leaf->data_offset = ndata_offset; 2336 hammer_modify_buffer_noundo(trans, data_buffer); 2337 2338 switch(mode) { 2339 case HAMMER_CREATE_MODE_UMIRROR: 2340 error = copyin(udata, ndata, leaf->data_len); 2341 if (error == 0) { 2342 if (hammer_crc_test_leaf(ndata, leaf) == 0) { 2343 hdkprintf("CRC DATA @ %016jx/%d MISMATCH ON PIPE\n", 2344 (intmax_t)ndata_offset, 2345 leaf->data_len); 2346 error = EINVAL; 2347 } else { 2348 error = hammer_cursor_localize_data( 2349 ndata, leaf); 2350 } 2351 } 2352 break; 2353 case HAMMER_CREATE_MODE_SYS: 2354 bcopy(udata, ndata, leaf->data_len); 2355 error = 0; 2356 hammer_crc_set_leaf(ndata, leaf); 2357 break; 2358 default: 2359 hpanic("bad mode %d", mode); 2360 break; /* NOT REACHED */ 2361 } 2362 hammer_modify_buffer_done(data_buffer); 2363 } else { 2364 leaf->data_offset = 0; 2365 error = 0; 2366 ndata = NULL; 2367 } 2368 if (error) 2369 goto failed; 2370 2371 /* 2372 * Do the insertion. This can fail with a EDEADLK or EALREADY 2373 */ 2374 cursor->flags |= HAMMER_CURSOR_INSERT; 2375 error = hammer_btree_lookup(cursor); 2376 if (error != ENOENT) { 2377 if (error == 0) 2378 error = EALREADY; 2379 goto failed; 2380 } 2381 error = hammer_btree_insert(cursor, leaf, &doprop); 2382 2383 /* 2384 * Cursor is left on current element, we want to skip it now. 2385 * (in case the caller is scanning) 2386 */ 2387 cursor->flags |= HAMMER_CURSOR_ATEDISK; 2388 cursor->flags &= ~HAMMER_CURSOR_INSERT; 2389 2390 /* 2391 * If the insertion happens to be creating (and not just replacing) 2392 * an inode we have to track it. 2393 */ 2394 if (error == 0 && 2395 leaf->base.rec_type == HAMMER_RECTYPE_INODE && 2396 leaf->base.delete_tid == 0) { 2397 hammer_modify_volume_field(trans, trans->rootvol, 2398 vol0_stat_inodes); 2399 ++trans->hmp->rootvol->ondisk->vol0_stat_inodes; 2400 hammer_modify_volume_done(trans->rootvol); 2401 } 2402 2403 /* 2404 * vol0_next_tid must track the highest TID stored in the filesystem. 2405 * We do not need to generate undo for this update. 2406 */ 2407 high_tid = leaf->base.create_tid; 2408 if (high_tid < leaf->base.delete_tid) 2409 high_tid = leaf->base.delete_tid; 2410 if (trans->rootvol->ondisk->vol0_next_tid < high_tid) { 2411 hammer_modify_volume_noundo(trans, trans->rootvol); 2412 trans->rootvol->ondisk->vol0_next_tid = high_tid; 2413 hammer_modify_volume_done(trans->rootvol); 2414 } 2415 2416 /* 2417 * WARNING! cursor's leaf pointer may have changed after 2418 * do_propagation returns. 2419 */ 2420 if (error == 0 && doprop) 2421 hammer_btree_do_propagation(cursor, leaf); 2422 2423 failed: 2424 /* 2425 * Cleanup 2426 */ 2427 if (error && leaf->data_offset) { 2428 hammer_blockmap_free(trans, leaf->data_offset, leaf->data_len); 2429 2430 } 2431 hammer_sync_unlock(trans); 2432 if (data_buffer) 2433 hammer_rel_buffer(data_buffer, 0); 2434 return (error); 2435 } 2436 2437 /* 2438 * Delete the B-Tree element at the current cursor and do any necessary 2439 * mirror propagation. 2440 * 2441 * The cursor must be properly positioned for an iteration on return but 2442 * may be pointing at an internal element. 2443 * 2444 * An element can be un-deleted by passing a delete_tid of 0 with 2445 * HAMMER_DELETE_ADJUST. 2446 * 2447 * This function will store the number of bytes deleted in *stat_bytes 2448 * if stat_bytes is not NULL. 2449 */ 2450 int 2451 hammer_delete_at_cursor(hammer_cursor_t cursor, int delete_flags, 2452 hammer_tid_t delete_tid, uint32_t delete_ts, 2453 int track, int64_t *stat_bytes) 2454 { 2455 struct hammer_btree_leaf_elm save_leaf; 2456 hammer_transaction_t trans; 2457 hammer_btree_leaf_elm_t leaf; 2458 hammer_node_t node; 2459 hammer_btree_elm_t elm; 2460 hammer_off_t data_offset; 2461 int32_t data_len; 2462 int64_t bytes; 2463 int ndelete; 2464 int error; 2465 int icount; 2466 int doprop; 2467 2468 error = hammer_cursor_upgrade(cursor); 2469 if (error) 2470 return(error); 2471 2472 trans = cursor->trans; 2473 node = cursor->node; 2474 elm = &node->ondisk->elms[cursor->index]; 2475 leaf = &elm->leaf; 2476 KKASSERT(elm->base.btype == HAMMER_BTREE_TYPE_RECORD); 2477 2478 hammer_sync_lock_sh(trans); 2479 bytes = 0; 2480 doprop = 0; 2481 icount = 0; 2482 2483 /* 2484 * Adjust the delete_tid. Update the mirror_tid propagation field 2485 * as well. delete_tid can be 0 (undelete -- used by mirroring). 2486 */ 2487 if (delete_flags & HAMMER_DELETE_ADJUST) { 2488 if (elm->base.rec_type == HAMMER_RECTYPE_INODE) { 2489 if (elm->leaf.base.delete_tid == 0 && delete_tid) 2490 icount = -1; 2491 if (elm->leaf.base.delete_tid && delete_tid == 0) 2492 icount = 1; 2493 } 2494 2495 hammer_modify_node(trans, node, elm, sizeof(*elm)); 2496 elm->leaf.base.delete_tid = delete_tid; 2497 elm->leaf.delete_ts = delete_ts; 2498 hammer_modify_node_done(node); 2499 2500 if (elm->leaf.base.delete_tid > node->ondisk->mirror_tid) { 2501 hammer_modify_node_field(trans, node, mirror_tid); 2502 node->ondisk->mirror_tid = elm->leaf.base.delete_tid; 2503 hammer_modify_node_done(node); 2504 doprop = 1; 2505 if (hammer_debug_general & 0x0002) { 2506 hdkprintf("propagate %016jx @%016jx\n", 2507 (intmax_t)elm->leaf.base.delete_tid, 2508 (intmax_t)node->node_offset); 2509 } 2510 } 2511 2512 /* 2513 * Adjust for the iteration. We have deleted the current 2514 * element and want to clear ATEDISK so the iteration does 2515 * not skip the element after, which now becomes the current 2516 * element. This element must be re-tested if doing an 2517 * iteration, which is handled by the RETEST flag. 2518 */ 2519 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 2520 cursor->flags |= HAMMER_CURSOR_RETEST; 2521 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 2522 } 2523 2524 /* 2525 * An on-disk record cannot have the same delete_tid 2526 * as its create_tid. In a chain of record updates 2527 * this could result in a duplicate record. 2528 */ 2529 KKASSERT(elm->leaf.base.delete_tid != 2530 elm->leaf.base.create_tid); 2531 } 2532 2533 /* 2534 * Destroy the B-Tree element if asked (typically if a nohistory 2535 * file or mount, or when called by the pruning code). 2536 * 2537 * Adjust the ATEDISK flag to properly support iterations. 2538 */ 2539 if (delete_flags & HAMMER_DELETE_DESTROY) { 2540 data_offset = elm->leaf.data_offset; 2541 data_len = elm->leaf.data_len; 2542 if (doprop) { 2543 save_leaf = elm->leaf; 2544 leaf = &save_leaf; 2545 } 2546 if (elm->base.rec_type == HAMMER_RECTYPE_INODE && 2547 elm->leaf.base.delete_tid == 0) { 2548 icount = -1; 2549 } 2550 2551 error = hammer_btree_delete(cursor, &ndelete); 2552 if (error == 0) { 2553 /* 2554 * The deletion moves the next element (if any) to 2555 * the current element position. We must clear 2556 * ATEDISK so this element is not skipped and we 2557 * must set RETEST to force any iteration to re-test 2558 * the element. 2559 */ 2560 if ((cursor->flags & HAMMER_CURSOR_DISKEOF) == 0) { 2561 cursor->flags |= HAMMER_CURSOR_RETEST; 2562 cursor->flags &= ~HAMMER_CURSOR_ATEDISK; 2563 } 2564 bytes += (ndelete * sizeof(struct hammer_node_ondisk)); 2565 2566 switch(HAMMER_ZONE(data_offset)) { 2567 case HAMMER_ZONE_LARGE_DATA: 2568 case HAMMER_ZONE_SMALL_DATA: 2569 case HAMMER_ZONE_META: 2570 hammer_blockmap_free(trans, 2571 data_offset, data_len); 2572 bytes += data_len; 2573 break; 2574 default: 2575 break; 2576 } 2577 } 2578 } 2579 2580 /* 2581 * Track inode count and next_tid. This is used by the mirroring 2582 * and PFS code. icount can be negative, zero, or positive. 2583 */ 2584 if (error == 0 && track) { 2585 if (icount) { 2586 hammer_modify_volume_field(trans, trans->rootvol, 2587 vol0_stat_inodes); 2588 trans->rootvol->ondisk->vol0_stat_inodes += icount; 2589 hammer_modify_volume_done(trans->rootvol); 2590 } 2591 if (trans->rootvol->ondisk->vol0_next_tid < delete_tid) { 2592 hammer_modify_volume_noundo(trans, trans->rootvol); 2593 trans->rootvol->ondisk->vol0_next_tid = delete_tid; 2594 hammer_modify_volume_done(trans->rootvol); 2595 } 2596 } 2597 2598 /* 2599 * mirror_tid propagation occurs if the node's mirror_tid had to be 2600 * updated while adjusting the delete_tid. 2601 * 2602 * This occurs when deleting even in nohistory mode, but does not 2603 * occur when pruning an already-deleted node. 2604 * 2605 * cursor->ip is NULL when called from the pruning, mirroring, 2606 * and pfs code. If non-NULL propagation will be conditionalized 2607 * on whether the PFS is in no-history mode or not. 2608 * 2609 * WARNING: cursor's leaf pointer may have changed after do_propagation 2610 * returns! 2611 */ 2612 if (doprop) { 2613 if (cursor->ip) 2614 hammer_btree_do_propagation(cursor, leaf); 2615 else 2616 hammer_btree_do_propagation(cursor, leaf); 2617 } 2618 if (stat_bytes) 2619 *stat_bytes = bytes; 2620 hammer_sync_unlock(trans); 2621 return (error); 2622 } 2623 2624 /* 2625 * Determine whether we can remove a directory. This routine checks whether 2626 * a directory is empty or not and enforces flush connectivity. 2627 * 2628 * Flush connectivity requires that we block if the target directory is 2629 * currently flushing, otherwise it may not end up in the same flush group. 2630 * 2631 * Returns 0 on success, ENOTEMPTY or EDEADLK (or other errors) on failure. 2632 */ 2633 int 2634 hammer_ip_check_directory_empty(hammer_transaction_t trans, hammer_inode_t ip) 2635 { 2636 struct hammer_cursor cursor; 2637 int error; 2638 2639 /* 2640 * Check directory empty 2641 */ 2642 hammer_init_cursor(trans, &cursor, &ip->cache[1], ip); 2643 2644 cursor.key_beg.localization = ip->obj_localization | 2645 hammer_dir_localization(ip); 2646 cursor.key_beg.obj_id = ip->obj_id; 2647 cursor.key_beg.create_tid = 0; 2648 cursor.key_beg.delete_tid = 0; 2649 cursor.key_beg.obj_type = 0; 2650 cursor.key_beg.rec_type = HAMMER_RECTYPE_ENTRY_START; 2651 cursor.key_beg.key = HAMMER_MIN_KEY; 2652 2653 cursor.key_end = cursor.key_beg; 2654 cursor.key_end.rec_type = HAMMER_RECTYPE_MAX; 2655 cursor.key_end.key = HAMMER_MAX_KEY; 2656 2657 cursor.asof = ip->obj_asof; 2658 cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF; 2659 2660 error = hammer_ip_first(&cursor); 2661 if (error == ENOENT) 2662 error = 0; 2663 else if (error == 0) 2664 error = ENOTEMPTY; 2665 hammer_done_cursor(&cursor); 2666 return(error); 2667 } 2668 2669 /* 2670 * Localize the data payload. Directory entries may need their 2671 * localization adjusted. 2672 */ 2673 static 2674 int 2675 hammer_cursor_localize_data(hammer_data_ondisk_t data, 2676 hammer_btree_leaf_elm_t leaf) 2677 { 2678 uint32_t localization; 2679 2680 if (leaf->base.rec_type == HAMMER_RECTYPE_DIRENTRY) { 2681 localization = leaf->base.localization & 2682 HAMMER_LOCALIZE_PSEUDOFS_MASK; 2683 if (data->entry.localization != localization) { 2684 data->entry.localization = localization; 2685 hammer_crc_set_leaf(data, leaf); 2686 } 2687 } 2688 return(0); 2689 } 2690