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