1 /* 2 * Copyright (c) 2011-2020 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@dragonflybsd.org> 6 * and Venkatesh Srinivas <vsrinivas@dragonflybsd.org> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 /* 36 * This subsystem implements most of the core support functions for 37 * the hammer2_chain structure. 38 * 39 * Chains are the in-memory version on media objects (volume header, inodes, 40 * indirect blocks, data blocks, etc). Chains represent a portion of the 41 * HAMMER2 topology. 42 * 43 * Chains are no-longer delete-duplicated. Instead, the original in-memory 44 * chain will be moved along with its block reference (e.g. for things like 45 * renames, hardlink operations, modifications, etc), and will be indexed 46 * on a secondary list for flush handling instead of propagating a flag 47 * upward to the root. 48 * 49 * Concurrent front-end operations can still run against backend flushes 50 * as long as they do not cross the current flush boundary. An operation 51 * running above the current flush (in areas not yet flushed) can become 52 * part of the current flush while ano peration running below the current 53 * flush can become part of the next flush. 54 */ 55 #include <sys/cdefs.h> 56 #include <sys/param.h> 57 #include <sys/systm.h> 58 #include <sys/types.h> 59 #include <sys/lock.h> 60 #include <sys/buf.h> 61 62 #include <crypto/sha2/sha2.h> 63 64 #include "hammer2.h" 65 66 static hammer2_chain_t *hammer2_chain_create_indirect( 67 hammer2_chain_t *parent, 68 hammer2_key_t key, int keybits, 69 hammer2_tid_t mtid, int for_type, int *errorp); 70 static int hammer2_chain_delete_obref(hammer2_chain_t *parent, 71 hammer2_chain_t *chain, 72 hammer2_tid_t mtid, int flags, 73 hammer2_blockref_t *obref); 74 static hammer2_chain_t *hammer2_combined_find( 75 hammer2_chain_t *parent, 76 hammer2_blockref_t *base, int count, 77 hammer2_key_t *key_nextp, 78 hammer2_key_t key_beg, hammer2_key_t key_end, 79 hammer2_blockref_t **brefp); 80 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain, 81 int depth); 82 static void hammer2_chain_lru_flush(hammer2_pfs_t *pmp); 83 84 /* 85 * There are many degenerate situations where an extreme rate of console 86 * output can occur from warnings and errors. Make sure this output does 87 * not impede operations. 88 */ 89 static struct krate krate_h2chk = { .freq = 5 }; 90 static struct krate krate_h2me = { .freq = 1 }; 91 static struct krate krate_h2em = { .freq = 1 }; 92 93 /* 94 * Basic RBTree for chains (core.rbtree). 95 */ 96 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp); 97 98 int 99 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2) 100 { 101 hammer2_key_t c1_beg; 102 hammer2_key_t c1_end; 103 hammer2_key_t c2_beg; 104 hammer2_key_t c2_end; 105 106 /* 107 * Compare chains. Overlaps are not supposed to happen and catch 108 * any software issues early we count overlaps as a match. 109 */ 110 c1_beg = chain1->bref.key; 111 c1_end = c1_beg + ((hammer2_key_t)1 << chain1->bref.keybits) - 1; 112 c2_beg = chain2->bref.key; 113 c2_end = c2_beg + ((hammer2_key_t)1 << chain2->bref.keybits) - 1; 114 115 if (c1_end < c2_beg) /* fully to the left */ 116 return(-1); 117 if (c1_beg > c2_end) /* fully to the right */ 118 return(1); 119 return(0); /* overlap (must not cross edge boundary) */ 120 } 121 122 /* 123 * Assert that a chain has no media data associated with it. 124 */ 125 static __inline void 126 hammer2_chain_assert_no_data(hammer2_chain_t *chain) 127 { 128 KKASSERT(chain->dio == NULL); 129 if (chain->bref.type != HAMMER2_BREF_TYPE_VOLUME && 130 chain->bref.type != HAMMER2_BREF_TYPE_FREEMAP && 131 chain->data) { 132 panic("hammer2_chain_assert_no_data: chain %p still has data", 133 chain); 134 } 135 } 136 137 /* 138 * Make a chain visible to the flusher. The flusher operates using a top-down 139 * recursion based on the ONFLUSH flag. It locates MODIFIED and UPDATE chains, 140 * flushes them, and updates blocks back to the volume root. 141 * 142 * This routine sets the ONFLUSH flag upward from the triggering chain until 143 * it hits an inode root or the volume root. Inode chains serve as inflection 144 * points, requiring the flusher to bridge across trees. Inodes include 145 * regular inodes, PFS roots (pmp->iroot), and the media super root 146 * (spmp->iroot). 147 */ 148 void 149 hammer2_chain_setflush(hammer2_chain_t *chain) 150 { 151 hammer2_chain_t *parent; 152 153 if ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) { 154 hammer2_spin_sh(&chain->core.spin); 155 while ((chain->flags & HAMMER2_CHAIN_ONFLUSH) == 0) { 156 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONFLUSH); 157 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) 158 break; 159 if ((parent = chain->parent) == NULL) 160 break; 161 hammer2_spin_sh(&parent->core.spin); 162 hammer2_spin_unsh(&chain->core.spin); 163 chain = parent; 164 } 165 hammer2_spin_unsh(&chain->core.spin); 166 } 167 } 168 169 /* 170 * Allocate a new disconnected chain element representing the specified 171 * bref. chain->refs is set to 1 and the passed bref is copied to 172 * chain->bref. chain->bytes is derived from the bref. 173 * 174 * chain->pmp inherits pmp unless the chain is an inode (other than the 175 * super-root inode). 176 * 177 * NOTE: Returns a referenced but unlocked (because there is no core) chain. 178 */ 179 hammer2_chain_t * 180 hammer2_chain_alloc(hammer2_dev_t *hmp, hammer2_pfs_t *pmp, 181 hammer2_blockref_t *bref) 182 { 183 hammer2_chain_t *chain; 184 u_int bytes; 185 186 /* 187 * Special case - radix of 0 indicates a chain that does not 188 * need a data reference (context is completely embedded in the 189 * bref). 190 */ 191 if ((int)(bref->data_off & HAMMER2_OFF_MASK_RADIX)) 192 bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX); 193 else 194 bytes = 0; 195 196 switch(bref->type) { 197 case HAMMER2_BREF_TYPE_INODE: 198 case HAMMER2_BREF_TYPE_INDIRECT: 199 case HAMMER2_BREF_TYPE_DATA: 200 case HAMMER2_BREF_TYPE_DIRENT: 201 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 202 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 203 case HAMMER2_BREF_TYPE_FREEMAP: 204 case HAMMER2_BREF_TYPE_VOLUME: 205 chain = kmalloc_obj(sizeof(*chain), hmp->mchain, 206 M_WAITOK | M_ZERO); 207 atomic_add_long(&hammer2_chain_allocs, 1); 208 break; 209 case HAMMER2_BREF_TYPE_EMPTY: 210 default: 211 panic("hammer2_chain_alloc: unrecognized blockref type: %d", 212 bref->type); 213 break; 214 } 215 216 /* 217 * Initialize the new chain structure. pmp must be set to NULL for 218 * chains belonging to the super-root topology of a device mount. 219 */ 220 if (pmp == hmp->spmp) 221 chain->pmp = NULL; 222 else 223 chain->pmp = pmp; 224 225 chain->hmp = hmp; 226 chain->bref = *bref; 227 chain->bytes = bytes; 228 chain->refs = 1; 229 chain->flags = HAMMER2_CHAIN_ALLOCATED; 230 lockinit(&chain->diolk, "chdio", 0, 0); 231 232 /* 233 * Set the PFS boundary flag if this chain represents a PFS root. 234 */ 235 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT) 236 atomic_set_int(&chain->flags, HAMMER2_CHAIN_PFSBOUNDARY); 237 hammer2_chain_core_init(chain); 238 239 return (chain); 240 } 241 242 /* 243 * Initialize a chain's core structure. This structure used to be allocated 244 * but is now embedded. 245 * 246 * The core is not locked. No additional refs on the chain are made. 247 */ 248 void 249 hammer2_chain_core_init(hammer2_chain_t *chain) 250 { 251 RB_INIT(&chain->core.rbtree); /* live chains */ 252 hammer2_mtx_init(&chain->lock, "h2chain"); 253 hammer2_spin_init(&chain->core.spin, "h2chain"); 254 } 255 256 /* 257 * Add a reference to a chain element, preventing its destruction. 258 * 259 * (can be called with spinlock held) 260 */ 261 void 262 hammer2_chain_ref(hammer2_chain_t *chain) 263 { 264 if (atomic_fetchadd_int(&chain->refs, 1) == 0) { 265 /* 266 * Just flag that the chain was used and should be recycled 267 * on the LRU if it encounters it later. 268 */ 269 if (chain->flags & HAMMER2_CHAIN_ONLRU) 270 atomic_set_int(&chain->flags, HAMMER2_CHAIN_LRUHINT); 271 272 #if 0 273 /* 274 * REMOVED - reduces contention, lru_list is more heuristical 275 * now. 276 * 277 * 0->non-zero transition must ensure that chain is removed 278 * from the LRU list. 279 * 280 * NOTE: Already holding lru_spin here so we cannot call 281 * hammer2_chain_ref() to get it off lru_list, do 282 * it manually. 283 */ 284 if (chain->flags & HAMMER2_CHAIN_ONLRU) { 285 hammer2_pfs_t *pmp = chain->pmp; 286 hammer2_spin_ex(&pmp->lru_spin); 287 if (chain->flags & HAMMER2_CHAIN_ONLRU) { 288 atomic_add_int(&pmp->lru_count, -1); 289 atomic_clear_int(&chain->flags, 290 HAMMER2_CHAIN_ONLRU); 291 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 292 } 293 hammer2_spin_unex(&pmp->lru_spin); 294 } 295 #endif 296 } 297 } 298 299 /* 300 * Ref a locked chain and force the data to be held across an unlock. 301 * Chain must be currently locked. The user of the chain who desires 302 * to release the hold must call hammer2_chain_lock_unhold() to relock 303 * and unhold the chain, then unlock normally, or may simply call 304 * hammer2_chain_drop_unhold() (which is safer against deadlocks). 305 */ 306 void 307 hammer2_chain_ref_hold(hammer2_chain_t *chain) 308 { 309 atomic_add_int(&chain->lockcnt, 1); 310 hammer2_chain_ref(chain); 311 } 312 313 /* 314 * Insert the chain in the core rbtree. 315 * 316 * Normal insertions are placed in the live rbtree. Insertion of a deleted 317 * chain is a special case used by the flush code that is placed on the 318 * unstaged deleted list to avoid confusing the live view. 319 */ 320 #define HAMMER2_CHAIN_INSERT_SPIN 0x0001 321 #define HAMMER2_CHAIN_INSERT_LIVE 0x0002 322 #define HAMMER2_CHAIN_INSERT_RACE 0x0004 323 324 static 325 int 326 hammer2_chain_insert(hammer2_chain_t *parent, hammer2_chain_t *chain, 327 int flags, int generation) 328 { 329 hammer2_chain_t *xchain; 330 int error = 0; 331 332 if (flags & HAMMER2_CHAIN_INSERT_SPIN) 333 hammer2_spin_ex(&parent->core.spin); 334 335 /* 336 * Interlocked by spinlock, check for race 337 */ 338 if ((flags & HAMMER2_CHAIN_INSERT_RACE) && 339 parent->core.generation != generation) { 340 error = HAMMER2_ERROR_EAGAIN; 341 goto failed; 342 } 343 344 /* 345 * Insert chain 346 */ 347 xchain = RB_INSERT(hammer2_chain_tree, &parent->core.rbtree, chain); 348 KASSERT(xchain == NULL, 349 ("hammer2_chain_insert: collision %p %p (key=%016jx)", 350 chain, xchain, chain->bref.key)); 351 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 352 chain->parent = parent; 353 ++parent->core.chain_count; 354 ++parent->core.generation; /* XXX incs for _get() too, XXX */ 355 356 /* 357 * We have to keep track of the effective live-view blockref count 358 * so the create code knows when to push an indirect block. 359 */ 360 if (flags & HAMMER2_CHAIN_INSERT_LIVE) 361 atomic_add_int(&parent->core.live_count, 1); 362 failed: 363 if (flags & HAMMER2_CHAIN_INSERT_SPIN) 364 hammer2_spin_unex(&parent->core.spin); 365 return error; 366 } 367 368 /* 369 * Drop the caller's reference to the chain. When the ref count drops to 370 * zero this function will try to disassociate the chain from its parent and 371 * deallocate it, then recursely drop the parent using the implied ref 372 * from the chain's chain->parent. 373 * 374 * Nobody should own chain's mutex on the 1->0 transition, unless this drop 375 * races an acquisition by another cpu. Therefore we can loop if we are 376 * unable to acquire the mutex, and refs is unlikely to be 1 unless we again 377 * race against another drop. 378 */ 379 void 380 hammer2_chain_drop(hammer2_chain_t *chain) 381 { 382 u_int refs; 383 384 KKASSERT(chain->refs > 0); 385 386 while (chain) { 387 refs = chain->refs; 388 cpu_ccfence(); 389 KKASSERT(refs > 0); 390 391 if (refs == 1) { 392 if (hammer2_mtx_ex_try(&chain->lock) == 0) 393 chain = hammer2_chain_lastdrop(chain, 0); 394 /* retry the same chain, or chain from lastdrop */ 395 } else { 396 if (atomic_cmpset_int(&chain->refs, refs, refs - 1)) 397 break; 398 /* retry the same chain */ 399 } 400 cpu_pause(); 401 } 402 } 403 404 /* 405 * Unhold a held and probably not-locked chain, ensure that the data is 406 * dropped on the 1->0 transition of lockcnt by obtaining an exclusive 407 * lock and then simply unlocking the chain. 408 */ 409 void 410 hammer2_chain_unhold(hammer2_chain_t *chain) 411 { 412 u_int lockcnt; 413 int iter = 0; 414 415 for (;;) { 416 lockcnt = chain->lockcnt; 417 cpu_ccfence(); 418 if (lockcnt > 1) { 419 if (atomic_cmpset_int(&chain->lockcnt, 420 lockcnt, lockcnt - 1)) { 421 break; 422 } 423 } else if (hammer2_mtx_ex_try(&chain->lock) == 0) { 424 hammer2_chain_unlock(chain); 425 break; 426 } else { 427 /* 428 * This situation can easily occur on SMP due to 429 * the gap inbetween the 1->0 transition and the 430 * final unlock. We cannot safely block on the 431 * mutex because lockcnt might go above 1. 432 * 433 * XXX Sleep for one tick if it takes too long. 434 */ 435 if (++iter > 1000) { 436 if (iter > 1000 + hz) { 437 kprintf("hammer2: h2race1 %p\n", chain); 438 iter = 1000; 439 } 440 tsleep(&iter, 0, "h2race1", 1); 441 } 442 cpu_pause(); 443 } 444 } 445 } 446 447 void 448 hammer2_chain_drop_unhold(hammer2_chain_t *chain) 449 { 450 hammer2_chain_unhold(chain); 451 hammer2_chain_drop(chain); 452 } 453 454 void 455 hammer2_chain_rehold(hammer2_chain_t *chain) 456 { 457 hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED); 458 atomic_add_int(&chain->lockcnt, 1); 459 hammer2_chain_unlock(chain); 460 } 461 462 /* 463 * Handles the (potential) last drop of chain->refs from 1->0. Called with 464 * the mutex exclusively locked, refs == 1, and lockcnt 0. SMP races are 465 * possible against refs and lockcnt. We must dispose of the mutex on chain. 466 * 467 * This function returns an unlocked chain for recursive drop or NULL. It 468 * can return the same chain if it determines it has raced another ref. 469 * 470 * -- 471 * 472 * When two chains need to be recursively dropped we use the chain we 473 * would otherwise free to placehold the additional chain. It's a bit 474 * convoluted but we can't just recurse without potentially blowing out 475 * the kernel stack. 476 * 477 * The chain cannot be freed if it has any children. 478 * The chain cannot be freed if flagged MODIFIED unless we can dispose of it. 479 * The chain cannot be freed if flagged UPDATE unless we can dispose of it. 480 * Any dedup registration can remain intact. 481 * 482 * The core spinlock is allowed to nest child-to-parent (not parent-to-child). 483 */ 484 static 485 hammer2_chain_t * 486 hammer2_chain_lastdrop(hammer2_chain_t *chain, int depth) 487 { 488 hammer2_pfs_t *pmp; 489 hammer2_dev_t *hmp; 490 hammer2_chain_t *parent; 491 hammer2_chain_t *rdrop; 492 493 /* 494 * We need chain's spinlock to interlock the sub-tree test. 495 * We already have chain's mutex, protecting chain->parent. 496 * 497 * Remember that chain->refs can be in flux. 498 */ 499 hammer2_spin_ex(&chain->core.spin); 500 501 if (chain->parent != NULL) { 502 /* 503 * If the chain has a parent the UPDATE bit prevents scrapping 504 * as the chain is needed to properly flush the parent. Try 505 * to complete the 1->0 transition and return NULL. Retry 506 * (return chain) if we are unable to complete the 1->0 507 * transition, else return NULL (nothing more to do). 508 * 509 * If the chain has a parent the MODIFIED bit prevents 510 * scrapping. 511 * 512 * Chains with UPDATE/MODIFIED are *not* put on the LRU list! 513 */ 514 if (chain->flags & (HAMMER2_CHAIN_UPDATE | 515 HAMMER2_CHAIN_MODIFIED)) { 516 if (atomic_cmpset_int(&chain->refs, 1, 0)) { 517 hammer2_spin_unex(&chain->core.spin); 518 hammer2_chain_assert_no_data(chain); 519 hammer2_mtx_unlock(&chain->lock); 520 chain = NULL; 521 } else { 522 hammer2_spin_unex(&chain->core.spin); 523 hammer2_mtx_unlock(&chain->lock); 524 } 525 return (chain); 526 } 527 /* spinlock still held */ 528 } else if (chain->bref.type == HAMMER2_BREF_TYPE_VOLUME || 529 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP) { 530 /* 531 * Retain the static vchain and fchain. Clear bits that 532 * are not relevant. Do not clear the MODIFIED bit, 533 * and certainly do not put it on the delayed-flush queue. 534 */ 535 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 536 } else { 537 /* 538 * The chain has no parent and can be flagged for destruction. 539 * Since it has no parent, UPDATE can also be cleared. 540 */ 541 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY); 542 if (chain->flags & HAMMER2_CHAIN_UPDATE) 543 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 544 545 /* 546 * If the chain has children we must propagate the DESTROY 547 * flag downward and rip the disconnected topology apart. 548 * This is accomplished by calling hammer2_flush() on the 549 * chain. 550 * 551 * Any dedup is already handled by the underlying DIO, so 552 * we do not have to specifically flush it here. 553 */ 554 if (chain->core.chain_count) { 555 hammer2_spin_unex(&chain->core.spin); 556 hammer2_flush(chain, HAMMER2_FLUSH_TOP | 557 HAMMER2_FLUSH_ALL); 558 hammer2_mtx_unlock(&chain->lock); 559 560 return(chain); /* retry drop */ 561 } 562 563 /* 564 * Otherwise we can scrap the MODIFIED bit if it is set, 565 * and continue along the freeing path. 566 * 567 * Be sure to clean-out any dedup bits. Without a parent 568 * this chain will no longer be visible to the flush code. 569 * Easy check data_off to avoid the volume root. 570 */ 571 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 572 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 573 atomic_add_long(&hammer2_count_modified_chains, -1); 574 if (chain->pmp) 575 hammer2_pfs_memory_wakeup(chain->pmp, -1); 576 } 577 /* spinlock still held */ 578 } 579 580 /* spinlock still held */ 581 582 /* 583 * If any children exist we must leave the chain intact with refs == 0. 584 * They exist because chains are retained below us which have refs or 585 * may require flushing. 586 * 587 * Retry (return chain) if we fail to transition the refs to 0, else 588 * return NULL indication nothing more to do. 589 * 590 * Chains with children are NOT put on the LRU list. 591 */ 592 if (chain->core.chain_count) { 593 if (atomic_cmpset_int(&chain->refs, 1, 0)) { 594 hammer2_spin_unex(&chain->core.spin); 595 hammer2_chain_assert_no_data(chain); 596 hammer2_mtx_unlock(&chain->lock); 597 chain = NULL; 598 } else { 599 hammer2_spin_unex(&chain->core.spin); 600 hammer2_mtx_unlock(&chain->lock); 601 } 602 return (chain); 603 } 604 /* spinlock still held */ 605 /* no chains left under us */ 606 607 /* 608 * chain->core has no children left so no accessors can get to our 609 * chain from there. Now we have to lock the parent core to interlock 610 * remaining possible accessors that might bump chain's refs before 611 * we can safely drop chain's refs with intent to free the chain. 612 */ 613 hmp = chain->hmp; 614 pmp = chain->pmp; /* can be NULL */ 615 rdrop = NULL; 616 617 parent = chain->parent; 618 619 /* 620 * WARNING! chain's spin lock is still held here, and other spinlocks 621 * will be acquired and released in the code below. We 622 * cannot be making fancy procedure calls! 623 */ 624 625 /* 626 * We can cache the chain if it is associated with a pmp 627 * and not flagged as being destroyed or requesting a full 628 * release. In this situation the chain is not removed 629 * from its parent, i.e. it can still be looked up. 630 * 631 * We intentionally do not cache DATA chains because these 632 * were likely used to load data into the logical buffer cache 633 * and will not be accessed again for some time. 634 */ 635 if ((chain->flags & 636 (HAMMER2_CHAIN_DESTROY | HAMMER2_CHAIN_RELEASE)) == 0 && 637 chain->pmp && 638 chain->bref.type != HAMMER2_BREF_TYPE_DATA) { 639 if (parent) 640 hammer2_spin_ex(&parent->core.spin); 641 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) { 642 /* 643 * 1->0 transition failed, retry. Do not drop 644 * the chain's data yet! 645 */ 646 if (parent) 647 hammer2_spin_unex(&parent->core.spin); 648 hammer2_spin_unex(&chain->core.spin); 649 hammer2_mtx_unlock(&chain->lock); 650 651 return(chain); 652 } 653 654 /* 655 * Success 656 */ 657 hammer2_chain_assert_no_data(chain); 658 659 /* 660 * Make sure we are on the LRU list, clean up excessive 661 * LRU entries. We can only really drop one but there might 662 * be other entries that we can remove from the lru_list 663 * without dropping. 664 * 665 * NOTE: HAMMER2_CHAIN_ONLRU may only be safely set when 666 * chain->core.spin AND pmp->lru_spin are held, but 667 * can be safely cleared only holding pmp->lru_spin. 668 */ 669 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) { 670 hammer2_spin_ex(&pmp->lru_spin); 671 if ((chain->flags & HAMMER2_CHAIN_ONLRU) == 0) { 672 atomic_set_int(&chain->flags, 673 HAMMER2_CHAIN_ONLRU); 674 TAILQ_INSERT_TAIL(&pmp->lru_list, 675 chain, lru_node); 676 atomic_add_int(&pmp->lru_count, 1); 677 } 678 if (pmp->lru_count < HAMMER2_LRU_LIMIT) 679 depth = 1; /* disable lru_list flush */ 680 hammer2_spin_unex(&pmp->lru_spin); 681 } else { 682 /* disable lru flush */ 683 depth = 1; 684 } 685 686 if (parent) { 687 hammer2_spin_unex(&parent->core.spin); 688 parent = NULL; /* safety */ 689 } 690 hammer2_spin_unex(&chain->core.spin); 691 hammer2_mtx_unlock(&chain->lock); 692 693 /* 694 * lru_list hysteresis (see above for depth overrides). 695 * Note that depth also prevents excessive lastdrop recursion. 696 */ 697 if (depth == 0) 698 hammer2_chain_lru_flush(pmp); 699 700 return NULL; 701 /* NOT REACHED */ 702 } 703 704 /* 705 * Make sure we are not on the LRU list. 706 */ 707 if (chain->flags & HAMMER2_CHAIN_ONLRU) { 708 hammer2_spin_ex(&pmp->lru_spin); 709 if (chain->flags & HAMMER2_CHAIN_ONLRU) { 710 atomic_add_int(&pmp->lru_count, -1); 711 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 712 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 713 } 714 hammer2_spin_unex(&pmp->lru_spin); 715 } 716 717 /* 718 * Spinlock the parent and try to drop the last ref on chain. 719 * On success determine if we should dispose of the chain 720 * (remove the chain from its parent, etc). 721 * 722 * (normal core locks are top-down recursive but we define 723 * core spinlocks as bottom-up recursive, so this is safe). 724 */ 725 if (parent) { 726 hammer2_spin_ex(&parent->core.spin); 727 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) { 728 /* 729 * 1->0 transition failed, retry. 730 */ 731 hammer2_spin_unex(&parent->core.spin); 732 hammer2_spin_unex(&chain->core.spin); 733 hammer2_mtx_unlock(&chain->lock); 734 735 return(chain); 736 } 737 738 /* 739 * 1->0 transition successful, parent spin held to prevent 740 * new lookups, chain spinlock held to protect parent field. 741 * Remove chain from the parent. 742 * 743 * If the chain is being removed from the parent's btree but 744 * is not blkmapped, we have to adjust live_count downward. If 745 * it is blkmapped then the blockref is retained in the parent 746 * as is its associated live_count. This case can occur when 747 * a chain added to the topology is unable to flush and is 748 * then later deleted. 749 */ 750 if (chain->flags & HAMMER2_CHAIN_ONRBTREE) { 751 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) && 752 (chain->flags & HAMMER2_CHAIN_BLKMAPPED) == 0) { 753 atomic_add_int(&parent->core.live_count, -1); 754 } 755 RB_REMOVE(hammer2_chain_tree, 756 &parent->core.rbtree, chain); 757 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 758 --parent->core.chain_count; 759 chain->parent = NULL; 760 } 761 762 /* 763 * If our chain was the last chain in the parent's core the 764 * core is now empty and its parent might have to be 765 * re-dropped if it has 0 refs. 766 */ 767 if (parent->core.chain_count == 0) { 768 rdrop = parent; 769 atomic_add_int(&rdrop->refs, 1); 770 /* 771 if (atomic_cmpset_int(&rdrop->refs, 0, 1) == 0) 772 rdrop = NULL; 773 */ 774 } 775 hammer2_spin_unex(&parent->core.spin); 776 parent = NULL; /* safety */ 777 /* FALL THROUGH */ 778 } else { 779 /* 780 * No-parent case. 781 */ 782 if (atomic_cmpset_int(&chain->refs, 1, 0) == 0) { 783 /* 784 * 1->0 transition failed, retry. 785 */ 786 hammer2_spin_unex(&parent->core.spin); 787 hammer2_spin_unex(&chain->core.spin); 788 hammer2_mtx_unlock(&chain->lock); 789 790 return(chain); 791 } 792 } 793 794 /* 795 * Successful 1->0 transition, no parent, no children... no way for 796 * anyone to ref this chain any more. We can clean-up and free it. 797 * 798 * We still have the core spinlock, and core's chain_count is 0. 799 * Any parent spinlock is gone. 800 */ 801 hammer2_spin_unex(&chain->core.spin); 802 hammer2_chain_assert_no_data(chain); 803 hammer2_mtx_unlock(&chain->lock); 804 KKASSERT(RB_EMPTY(&chain->core.rbtree) && 805 chain->core.chain_count == 0); 806 807 /* 808 * All locks are gone, no pointers remain to the chain, finish 809 * freeing it. 810 */ 811 KKASSERT((chain->flags & (HAMMER2_CHAIN_UPDATE | 812 HAMMER2_CHAIN_MODIFIED)) == 0); 813 814 /* 815 * Once chain resources are gone we can use the now dead chain 816 * structure to placehold what might otherwise require a recursive 817 * drop, because we have potentially two things to drop and can only 818 * return one directly. 819 */ 820 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) { 821 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ALLOCATED); 822 chain->hmp = NULL; 823 kfree_obj(chain, hmp->mchain); 824 atomic_add_long(&hammer2_chain_allocs, -1); 825 } 826 827 /* 828 * Possible chaining loop when parent re-drop needed. 829 */ 830 return(rdrop); 831 } 832 833 /* 834 * Heuristical flush of the LRU, try to reduce the number of entries 835 * on the LRU to (HAMMER2_LRU_LIMIT * 2 / 3). This procedure is called 836 * only when lru_count exceeds HAMMER2_LRU_LIMIT. 837 */ 838 static 839 void 840 hammer2_chain_lru_flush(hammer2_pfs_t *pmp) 841 { 842 hammer2_chain_t *chain; 843 844 again: 845 chain = NULL; 846 hammer2_spin_ex(&pmp->lru_spin); 847 while (pmp->lru_count > HAMMER2_LRU_LIMIT * 2 / 3) { 848 /* 849 * Pick a chain off the lru_list, just recycle it quickly 850 * if LRUHINT is set (the chain was ref'd but left on 851 * the lru_list, so cycle to the end). 852 */ 853 chain = TAILQ_FIRST(&pmp->lru_list); 854 TAILQ_REMOVE(&pmp->lru_list, chain, lru_node); 855 856 if (chain->flags & HAMMER2_CHAIN_LRUHINT) { 857 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_LRUHINT); 858 TAILQ_INSERT_TAIL(&pmp->lru_list, chain, lru_node); 859 chain = NULL; 860 continue; 861 } 862 863 /* 864 * Ok, we are off the LRU. We must adjust refs before we 865 * can safely clear the ONLRU flag. 866 */ 867 atomic_add_int(&pmp->lru_count, -1); 868 if (atomic_cmpset_int(&chain->refs, 0, 1)) { 869 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 870 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE); 871 break; 872 } 873 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONLRU); 874 chain = NULL; 875 } 876 hammer2_spin_unex(&pmp->lru_spin); 877 if (chain == NULL) 878 return; 879 880 /* 881 * If we picked a chain off the lru list we may be able to lastdrop 882 * it. Use a depth of 1 to prevent excessive lastdrop recursion. 883 */ 884 while (chain) { 885 u_int refs; 886 887 refs = chain->refs; 888 cpu_ccfence(); 889 KKASSERT(refs > 0); 890 891 if (refs == 1) { 892 if (hammer2_mtx_ex_try(&chain->lock) == 0) 893 chain = hammer2_chain_lastdrop(chain, 1); 894 /* retry the same chain, or chain from lastdrop */ 895 } else { 896 if (atomic_cmpset_int(&chain->refs, refs, refs - 1)) 897 break; 898 /* retry the same chain */ 899 } 900 cpu_pause(); 901 } 902 goto again; 903 } 904 905 /* 906 * On last lock release. 907 */ 908 static hammer2_io_t * 909 hammer2_chain_drop_data(hammer2_chain_t *chain) 910 { 911 hammer2_io_t *dio; 912 913 if ((dio = chain->dio) != NULL) { 914 chain->dio = NULL; 915 chain->data = NULL; 916 } else { 917 switch(chain->bref.type) { 918 case HAMMER2_BREF_TYPE_VOLUME: 919 case HAMMER2_BREF_TYPE_FREEMAP: 920 break; 921 default: 922 if (chain->data != NULL) { 923 hammer2_spin_unex(&chain->core.spin); 924 panic("chain data not null: " 925 "chain %p bref %016jx.%02x " 926 "refs %d parent %p dio %p data %p", 927 chain, chain->bref.data_off, 928 chain->bref.type, chain->refs, 929 chain->parent, 930 chain->dio, chain->data); 931 } 932 KKASSERT(chain->data == NULL); 933 break; 934 } 935 } 936 return dio; 937 } 938 939 /* 940 * Lock a referenced chain element, acquiring its data with I/O if necessary, 941 * and specify how you would like the data to be resolved. 942 * 943 * If an I/O or other fatal error occurs, chain->error will be set to non-zero. 944 * 945 * The lock is allowed to recurse, multiple locking ops will aggregate 946 * the requested resolve types. Once data is assigned it will not be 947 * removed until the last unlock. 948 * 949 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element. 950 * (typically used to avoid device/logical buffer 951 * aliasing for data) 952 * 953 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in 954 * the INITIAL-create state (indirect blocks only). 955 * 956 * Do not resolve data elements for DATA chains. 957 * (typically used to avoid device/logical buffer 958 * aliasing for data) 959 * 960 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element. 961 * 962 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise 963 * it will be locked exclusive. 964 * 965 * HAMMER2_RESOLVE_NONBLOCK- (flag) The chain is locked non-blocking. If 966 * the lock fails, EAGAIN is returned. 967 * 968 * NOTE: Embedded elements (volume header, inodes) are always resolved 969 * regardless. 970 * 971 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded 972 * element will instantiate and zero its buffer, and flush it on 973 * release. 974 * 975 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE 976 * so as not to instantiate a device buffer, which could alias against 977 * a logical file buffer. However, if ALWAYS is specified the 978 * device buffer will be instantiated anyway. 979 * 980 * NOTE: The return value is always 0 unless NONBLOCK is specified, in which 981 * case it can be either 0 or EAGAIN. 982 * 983 * WARNING! This function blocks on I/O if data needs to be fetched. This 984 * blocking can run concurrent with other compatible lock holders 985 * who do not need data returning. The lock is not upgraded to 986 * exclusive during a data fetch, a separate bit is used to 987 * interlock I/O. However, an exclusive lock holder can still count 988 * on being interlocked against an I/O fetch managed by a shared 989 * lock holder. 990 */ 991 int 992 hammer2_chain_lock(hammer2_chain_t *chain, int how) 993 { 994 KKASSERT(chain->refs > 0); 995 996 if (how & HAMMER2_RESOLVE_NONBLOCK) { 997 /* 998 * We still have to bump lockcnt before acquiring the lock, 999 * even for non-blocking operation, because the unlock code 1000 * live-loops on lockcnt == 1 when dropping the last lock. 1001 * 1002 * If the non-blocking operation fails we have to use an 1003 * unhold sequence to undo the mess. 1004 * 1005 * NOTE: LOCKAGAIN must always succeed without blocking, 1006 * even if NONBLOCK is specified. 1007 */ 1008 atomic_add_int(&chain->lockcnt, 1); 1009 if (how & HAMMER2_RESOLVE_SHARED) { 1010 if (how & HAMMER2_RESOLVE_LOCKAGAIN) { 1011 hammer2_mtx_sh_again(&chain->lock); 1012 } else { 1013 if (hammer2_mtx_sh_try(&chain->lock) != 0) { 1014 hammer2_chain_unhold(chain); 1015 return EAGAIN; 1016 } 1017 } 1018 } else { 1019 if (hammer2_mtx_ex_try(&chain->lock) != 0) { 1020 hammer2_chain_unhold(chain); 1021 return EAGAIN; 1022 } 1023 } 1024 } else { 1025 /* 1026 * Get the appropriate lock. If LOCKAGAIN is flagged with 1027 * SHARED the caller expects a shared lock to already be 1028 * present and we are giving it another ref. This case must 1029 * importantly not block if there is a pending exclusive lock 1030 * request. 1031 */ 1032 atomic_add_int(&chain->lockcnt, 1); 1033 if (how & HAMMER2_RESOLVE_SHARED) { 1034 if (how & HAMMER2_RESOLVE_LOCKAGAIN) { 1035 hammer2_mtx_sh_again(&chain->lock); 1036 } else { 1037 hammer2_mtx_sh(&chain->lock); 1038 } 1039 } else { 1040 hammer2_mtx_ex(&chain->lock); 1041 } 1042 } 1043 1044 /* 1045 * If we already have a valid data pointer make sure the data is 1046 * synchronized to the current cpu, and then no further action is 1047 * necessary. 1048 */ 1049 if (chain->data) { 1050 if (chain->dio) 1051 hammer2_io_bkvasync(chain->dio); 1052 return 0; 1053 } 1054 1055 /* 1056 * Do we have to resolve the data? This is generally only 1057 * applicable to HAMMER2_BREF_TYPE_DATA which is special-cased. 1058 * Other BREF types expects the data to be there. 1059 */ 1060 switch(how & HAMMER2_RESOLVE_MASK) { 1061 case HAMMER2_RESOLVE_NEVER: 1062 return 0; 1063 case HAMMER2_RESOLVE_MAYBE: 1064 if (chain->flags & HAMMER2_CHAIN_INITIAL) 1065 return 0; 1066 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) 1067 return 0; 1068 #if 0 1069 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) 1070 return 0; 1071 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) 1072 return 0; 1073 #endif 1074 /* fall through */ 1075 case HAMMER2_RESOLVE_ALWAYS: 1076 default: 1077 break; 1078 } 1079 1080 /* 1081 * Caller requires data 1082 */ 1083 hammer2_chain_load_data(chain); 1084 1085 return 0; 1086 } 1087 1088 /* 1089 * Lock the chain, retain the hold, and drop the data persistence count. 1090 * The data should remain valid because we never transitioned lockcnt 1091 * through 0. 1092 */ 1093 void 1094 hammer2_chain_lock_unhold(hammer2_chain_t *chain, int how) 1095 { 1096 hammer2_chain_lock(chain, how); 1097 atomic_add_int(&chain->lockcnt, -1); 1098 } 1099 1100 #if 0 1101 /* 1102 * Downgrade an exclusive chain lock to a shared chain lock. 1103 * 1104 * NOTE: There is no upgrade equivalent due to the ease of 1105 * deadlocks in that direction. 1106 */ 1107 void 1108 hammer2_chain_lock_downgrade(hammer2_chain_t *chain) 1109 { 1110 hammer2_mtx_downgrade(&chain->lock); 1111 } 1112 #endif 1113 1114 /* 1115 * Issue I/O and install chain->data. Caller must hold a chain lock, lock 1116 * may be of any type. 1117 * 1118 * Once chain->data is set it cannot be disposed of until all locks are 1119 * released. 1120 * 1121 * Make sure the data is synchronized to the current cpu. 1122 */ 1123 void 1124 hammer2_chain_load_data(hammer2_chain_t *chain) 1125 { 1126 hammer2_blockref_t *bref; 1127 hammer2_dev_t *hmp; 1128 hammer2_io_t *dio; 1129 char *bdata; 1130 int error; 1131 1132 /* 1133 * Degenerate case, data already present, or chain has no media 1134 * reference to load. 1135 */ 1136 KKASSERT(chain->lock.mtx_lock & MTX_MASK); 1137 if (chain->data) { 1138 if (chain->dio) 1139 hammer2_io_bkvasync(chain->dio); 1140 return; 1141 } 1142 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0) 1143 return; 1144 1145 hmp = chain->hmp; 1146 KKASSERT(hmp != NULL); 1147 1148 /* 1149 * Gain the IOINPROG bit, interlocked block. 1150 */ 1151 for (;;) { 1152 u_int oflags; 1153 u_int nflags; 1154 1155 oflags = chain->flags; 1156 cpu_ccfence(); 1157 if (oflags & HAMMER2_CHAIN_IOINPROG) { 1158 nflags = oflags | HAMMER2_CHAIN_IOSIGNAL; 1159 tsleep_interlock(&chain->flags, 0); 1160 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) { 1161 tsleep(&chain->flags, PINTERLOCKED, 1162 "h2iocw", 0); 1163 } 1164 /* retry */ 1165 } else { 1166 nflags = oflags | HAMMER2_CHAIN_IOINPROG; 1167 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) { 1168 break; 1169 } 1170 /* retry */ 1171 } 1172 } 1173 1174 /* 1175 * We own CHAIN_IOINPROG 1176 * 1177 * Degenerate case if we raced another load. 1178 */ 1179 if (chain->data) { 1180 if (chain->dio) 1181 hammer2_io_bkvasync(chain->dio); 1182 goto done; 1183 } 1184 1185 /* 1186 * We must resolve to a device buffer, either by issuing I/O or 1187 * by creating a zero-fill element. We do not mark the buffer 1188 * dirty when creating a zero-fill element (the hammer2_chain_modify() 1189 * API must still be used to do that). 1190 * 1191 * The device buffer is variable-sized in powers of 2 down 1192 * to HAMMER2_MIN_ALLOC (typically 1K). A 64K physical storage 1193 * chunk always contains buffers of the same size. (XXX) 1194 * 1195 * The minimum physical IO size may be larger than the variable 1196 * block size. 1197 */ 1198 bref = &chain->bref; 1199 1200 /* 1201 * The getblk() optimization can only be used on newly created 1202 * elements if the physical block size matches the request. 1203 */ 1204 if (chain->flags & HAMMER2_CHAIN_INITIAL) { 1205 error = hammer2_io_new(hmp, bref->type, 1206 bref->data_off, chain->bytes, 1207 &chain->dio); 1208 } else { 1209 error = hammer2_io_bread(hmp, bref->type, 1210 bref->data_off, chain->bytes, 1211 &chain->dio); 1212 hammer2_adjreadcounter(chain->bref.type, chain->bytes); 1213 } 1214 if (error) { 1215 chain->error = HAMMER2_ERROR_EIO; 1216 kprintf("hammer2_chain_load_data: I/O error %016jx: %d\n", 1217 (intmax_t)bref->data_off, error); 1218 hammer2_io_bqrelse(&chain->dio); 1219 goto done; 1220 } 1221 chain->error = 0; 1222 1223 /* 1224 * This isn't perfect and can be ignored on OSs which do not have 1225 * an indication as to whether a buffer is coming from cache or 1226 * if I/O was actually issued for the read. TESTEDGOOD will work 1227 * pretty well without the B_IOISSUED logic because chains are 1228 * cached, but in that situation (without B_IOISSUED) it will not 1229 * detect whether a re-read via I/O is corrupted verses the original 1230 * read. 1231 * 1232 * We can't re-run the CRC on every fresh lock. That would be 1233 * insanely expensive. 1234 * 1235 * If the underlying kernel buffer covers the entire chain we can 1236 * use the B_IOISSUED indication to determine if we have to re-run 1237 * the CRC on chain data for chains that managed to stay cached 1238 * across the kernel disposal of the original buffer. 1239 */ 1240 if ((dio = chain->dio) != NULL && dio->bp) { 1241 struct buf *bp = dio->bp; 1242 1243 if (dio->psize == chain->bytes && 1244 (bp->b_flags & B_IOISSUED)) { 1245 atomic_clear_int(&chain->flags, 1246 HAMMER2_CHAIN_TESTEDGOOD); 1247 bp->b_flags &= ~B_IOISSUED; 1248 } 1249 } 1250 1251 /* 1252 * NOTE: A locked chain's data cannot be modified without first 1253 * calling hammer2_chain_modify(). 1254 */ 1255 1256 /* 1257 * NOTE: hammer2_io_data() call issues bkvasync() 1258 */ 1259 bdata = hammer2_io_data(chain->dio, chain->bref.data_off); 1260 1261 if (chain->flags & HAMMER2_CHAIN_INITIAL) { 1262 /* 1263 * Clear INITIAL. In this case we used io_new() and the 1264 * buffer has been zero'd and marked dirty. 1265 * 1266 * CHAIN_MODIFIED has not been set yet, and we leave it 1267 * that way for now. Set a temporary CHAIN_NOTTESTED flag 1268 * to prevent hammer2_chain_testcheck() from trying to match 1269 * a check code that has not yet been generated. This bit 1270 * should NOT end up on the actual media. 1271 */ 1272 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1273 atomic_set_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED); 1274 } else if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 1275 /* 1276 * check data not currently synchronized due to 1277 * modification. XXX assumes data stays in the buffer 1278 * cache, which might not be true (need biodep on flush 1279 * to calculate crc? or simple crc?). 1280 */ 1281 } else if ((chain->flags & HAMMER2_CHAIN_TESTEDGOOD) == 0) { 1282 if (hammer2_chain_testcheck(chain, bdata) == 0) { 1283 chain->error = HAMMER2_ERROR_CHECK; 1284 } else { 1285 atomic_set_int(&chain->flags, HAMMER2_CHAIN_TESTEDGOOD); 1286 } 1287 } 1288 1289 /* 1290 * Setup the data pointer, either pointing it to an embedded data 1291 * structure and copying the data from the buffer, or pointing it 1292 * into the buffer. 1293 * 1294 * The buffer is not retained when copying to an embedded data 1295 * structure in order to avoid potential deadlocks or recursions 1296 * on the same physical buffer. 1297 * 1298 * WARNING! Other threads can start using the data the instant we 1299 * set chain->data non-NULL. 1300 */ 1301 switch (bref->type) { 1302 case HAMMER2_BREF_TYPE_VOLUME: 1303 case HAMMER2_BREF_TYPE_FREEMAP: 1304 /* 1305 * Copy data from bp to embedded buffer 1306 */ 1307 panic("hammer2_chain_load_data: unresolved volume header"); 1308 break; 1309 case HAMMER2_BREF_TYPE_DIRENT: 1310 KKASSERT(chain->bytes != 0); 1311 /* fall through */ 1312 case HAMMER2_BREF_TYPE_INODE: 1313 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1314 case HAMMER2_BREF_TYPE_INDIRECT: 1315 case HAMMER2_BREF_TYPE_DATA: 1316 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1317 default: 1318 /* 1319 * Point data at the device buffer and leave dio intact. 1320 */ 1321 chain->data = (void *)bdata; 1322 break; 1323 } 1324 1325 /* 1326 * Release HAMMER2_CHAIN_IOINPROG and signal waiters if requested. 1327 */ 1328 done: 1329 for (;;) { 1330 u_int oflags; 1331 u_int nflags; 1332 1333 oflags = chain->flags; 1334 nflags = oflags & ~(HAMMER2_CHAIN_IOINPROG | 1335 HAMMER2_CHAIN_IOSIGNAL); 1336 KKASSERT(oflags & HAMMER2_CHAIN_IOINPROG); 1337 if (atomic_cmpset_int(&chain->flags, oflags, nflags)) { 1338 if (oflags & HAMMER2_CHAIN_IOSIGNAL) 1339 wakeup(&chain->flags); 1340 break; 1341 } 1342 } 1343 } 1344 1345 /* 1346 * Unlock and deref a chain element. 1347 * 1348 * Remember that the presence of children under chain prevent the chain's 1349 * destruction but do not add additional references, so the dio will still 1350 * be dropped. 1351 */ 1352 void 1353 hammer2_chain_unlock(hammer2_chain_t *chain) 1354 { 1355 hammer2_io_t *dio; 1356 u_int lockcnt; 1357 int iter = 0; 1358 1359 /* 1360 * If multiple locks are present (or being attempted) on this 1361 * particular chain we can just unlock, drop refs, and return. 1362 * 1363 * Otherwise fall-through on the 1->0 transition. 1364 */ 1365 for (;;) { 1366 lockcnt = chain->lockcnt; 1367 KKASSERT(lockcnt > 0); 1368 cpu_ccfence(); 1369 if (lockcnt > 1) { 1370 if (atomic_cmpset_int(&chain->lockcnt, 1371 lockcnt, lockcnt - 1)) { 1372 hammer2_mtx_unlock(&chain->lock); 1373 return; 1374 } 1375 } else if (hammer2_mtx_upgrade_try(&chain->lock) == 0) { 1376 /* while holding the mutex exclusively */ 1377 if (atomic_cmpset_int(&chain->lockcnt, 1, 0)) 1378 break; 1379 } else { 1380 /* 1381 * This situation can easily occur on SMP due to 1382 * the gap inbetween the 1->0 transition and the 1383 * final unlock. We cannot safely block on the 1384 * mutex because lockcnt might go above 1. 1385 * 1386 * XXX Sleep for one tick if it takes too long. 1387 */ 1388 if (++iter > 1000) { 1389 if (iter > 1000 + hz) { 1390 kprintf("hammer2: h2race2 %p\n", chain); 1391 iter = 1000; 1392 } 1393 tsleep(&iter, 0, "h2race2", 1); 1394 } 1395 cpu_pause(); 1396 } 1397 /* retry */ 1398 } 1399 1400 /* 1401 * Last unlock / mutex upgraded to exclusive. Drop the data 1402 * reference. 1403 */ 1404 dio = hammer2_chain_drop_data(chain); 1405 if (dio) 1406 hammer2_io_bqrelse(&dio); 1407 hammer2_mtx_unlock(&chain->lock); 1408 } 1409 1410 /* 1411 * Unlock and hold chain data intact 1412 */ 1413 void 1414 hammer2_chain_unlock_hold(hammer2_chain_t *chain) 1415 { 1416 atomic_add_int(&chain->lockcnt, 1); 1417 hammer2_chain_unlock(chain); 1418 } 1419 1420 /* 1421 * Helper to obtain the blockref[] array base and count for a chain. 1422 * 1423 * XXX Not widely used yet, various use cases need to be validated and 1424 * converted to use this function. 1425 */ 1426 static 1427 hammer2_blockref_t * 1428 hammer2_chain_base_and_count(hammer2_chain_t *parent, int *countp) 1429 { 1430 hammer2_blockref_t *base; 1431 int count; 1432 1433 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 1434 base = NULL; 1435 1436 switch(parent->bref.type) { 1437 case HAMMER2_BREF_TYPE_INODE: 1438 count = HAMMER2_SET_COUNT; 1439 break; 1440 case HAMMER2_BREF_TYPE_INDIRECT: 1441 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1442 count = parent->bytes / sizeof(hammer2_blockref_t); 1443 break; 1444 case HAMMER2_BREF_TYPE_VOLUME: 1445 count = HAMMER2_SET_COUNT; 1446 break; 1447 case HAMMER2_BREF_TYPE_FREEMAP: 1448 count = HAMMER2_SET_COUNT; 1449 break; 1450 default: 1451 panic("hammer2_chain_base_and_count: " 1452 "unrecognized blockref type: %d", 1453 parent->bref.type); 1454 count = 0; 1455 break; 1456 } 1457 } else { 1458 switch(parent->bref.type) { 1459 case HAMMER2_BREF_TYPE_INODE: 1460 base = &parent->data->ipdata.u.blockset.blockref[0]; 1461 count = HAMMER2_SET_COUNT; 1462 break; 1463 case HAMMER2_BREF_TYPE_INDIRECT: 1464 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1465 base = &parent->data->npdata[0]; 1466 count = parent->bytes / sizeof(hammer2_blockref_t); 1467 break; 1468 case HAMMER2_BREF_TYPE_VOLUME: 1469 base = &parent->data->voldata. 1470 sroot_blockset.blockref[0]; 1471 count = HAMMER2_SET_COUNT; 1472 break; 1473 case HAMMER2_BREF_TYPE_FREEMAP: 1474 base = &parent->data->blkset.blockref[0]; 1475 count = HAMMER2_SET_COUNT; 1476 break; 1477 default: 1478 panic("hammer2_chain_base_and_count: " 1479 "unrecognized blockref type: %d", 1480 parent->bref.type); 1481 base = NULL; 1482 count = 0; 1483 break; 1484 } 1485 } 1486 *countp = count; 1487 1488 return base; 1489 } 1490 1491 /* 1492 * This counts the number of live blockrefs in a block array and 1493 * also calculates the point at which all remaining blockrefs are empty. 1494 * This routine can only be called on a live chain. 1495 * 1496 * Caller holds the chain locked, but possibly with a shared lock. We 1497 * must use an exclusive spinlock to prevent corruption. 1498 * 1499 * NOTE: Flag is not set until after the count is complete, allowing 1500 * callers to test the flag without holding the spinlock. 1501 * 1502 * NOTE: If base is NULL the related chain is still in the INITIAL 1503 * state and there are no blockrefs to count. 1504 * 1505 * NOTE: live_count may already have some counts accumulated due to 1506 * creation and deletion and could even be initially negative. 1507 */ 1508 void 1509 hammer2_chain_countbrefs(hammer2_chain_t *chain, 1510 hammer2_blockref_t *base, int count) 1511 { 1512 hammer2_spin_ex(&chain->core.spin); 1513 if ((chain->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) { 1514 if (base) { 1515 while (--count >= 0) { 1516 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY) 1517 break; 1518 } 1519 chain->core.live_zero = count + 1; 1520 while (count >= 0) { 1521 if (base[count].type != HAMMER2_BREF_TYPE_EMPTY) 1522 atomic_add_int(&chain->core.live_count, 1523 1); 1524 --count; 1525 } 1526 } else { 1527 chain->core.live_zero = 0; 1528 } 1529 /* else do not modify live_count */ 1530 atomic_set_int(&chain->flags, HAMMER2_CHAIN_COUNTEDBREFS); 1531 } 1532 hammer2_spin_unex(&chain->core.spin); 1533 } 1534 1535 /* 1536 * Resize the chain's physical storage allocation in-place. This function does 1537 * not usually adjust the data pointer and must be followed by (typically) a 1538 * hammer2_chain_modify() call to copy any old data over and adjust the 1539 * data pointer. 1540 * 1541 * Chains can be resized smaller without reallocating the storage. Resizing 1542 * larger will reallocate the storage. Excess or prior storage is reclaimed 1543 * asynchronously at a later time. 1544 * 1545 * An nradix value of 0 is special-cased to mean that the storage should 1546 * be disassociated, that is the chain is being resized to 0 bytes (not 1 1547 * byte). 1548 * 1549 * Must be passed an exclusively locked parent and chain. 1550 * 1551 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order 1552 * to avoid instantiating a device buffer that conflicts with the vnode data 1553 * buffer. However, because H2 can compress or encrypt data, the chain may 1554 * have a dio assigned to it in those situations, and they do not conflict. 1555 * 1556 * XXX return error if cannot resize. 1557 */ 1558 int 1559 hammer2_chain_resize(hammer2_chain_t *chain, 1560 hammer2_tid_t mtid, hammer2_off_t dedup_off, 1561 int nradix, int flags) 1562 { 1563 hammer2_dev_t *hmp; 1564 size_t obytes; 1565 size_t nbytes; 1566 int error; 1567 1568 hmp = chain->hmp; 1569 1570 /* 1571 * Only data and indirect blocks can be resized for now. 1572 * (The volu root, inodes, and freemap elements use a fixed size). 1573 */ 1574 KKASSERT(chain != &hmp->vchain); 1575 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA || 1576 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 1577 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT); 1578 1579 /* 1580 * Nothing to do if the element is already the proper size 1581 */ 1582 obytes = chain->bytes; 1583 nbytes = (nradix) ? (1U << nradix) : 0; 1584 if (obytes == nbytes) 1585 return (chain->error); 1586 1587 /* 1588 * Make sure the old data is instantiated so we can copy it. If this 1589 * is a data block, the device data may be superfluous since the data 1590 * might be in a logical block, but compressed or encrypted data is 1591 * another matter. 1592 * 1593 * NOTE: The modify will set BLKMAPUPD for us if BLKMAPPED is set. 1594 */ 1595 error = hammer2_chain_modify(chain, mtid, dedup_off, 0); 1596 if (error) 1597 return error; 1598 1599 /* 1600 * Reallocate the block, even if making it smaller (because different 1601 * block sizes may be in different regions). 1602 * 1603 * NOTE: Operation does not copy the data and may only be used 1604 * to resize data blocks in-place, or directory entry blocks 1605 * which are about to be modified in some manner. 1606 */ 1607 error = hammer2_freemap_alloc(chain, nbytes); 1608 if (error) 1609 return error; 1610 1611 chain->bytes = nbytes; 1612 1613 /* 1614 * We don't want the followup chain_modify() to try to copy data 1615 * from the old (wrong-sized) buffer. It won't know how much to 1616 * copy. This case should only occur during writes when the 1617 * originator already has the data to write in-hand. 1618 */ 1619 if (chain->dio) { 1620 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA || 1621 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT); 1622 hammer2_io_brelse(&chain->dio); 1623 chain->data = NULL; 1624 } 1625 return (chain->error); 1626 } 1627 1628 /* 1629 * Set the chain modified so its data can be changed by the caller, or 1630 * install deduplicated data. The caller must call this routine for each 1631 * set of modifications it makes, even if the chain is already flagged 1632 * MODIFIED. 1633 * 1634 * Sets bref.modify_tid to mtid only if mtid != 0. Note that bref.modify_tid 1635 * is a CLC (cluster level change) field and is not updated by parent 1636 * propagation during a flush. 1637 * 1638 * Returns an appropriate HAMMER2_ERROR_* code, which will generally reflect 1639 * chain->error except for HAMMER2_ERROR_ENOSPC. If the allocation fails 1640 * due to no space available, HAMMER2_ERROR_ENOSPC is returned and the chain 1641 * remains unmodified with its old data ref intact and chain->error 1642 * unchanged. 1643 * 1644 * Dedup Handling 1645 * 1646 * If the DEDUPABLE flag is set in the chain the storage must be reallocated 1647 * even if the chain is still flagged MODIFIED. In this case the chain's 1648 * DEDUPABLE flag will be cleared once the new storage has been assigned. 1649 * 1650 * If the caller passes a non-zero dedup_off we will use it to assign the 1651 * new storage. The MODIFIED flag will be *CLEARED* in this case, and 1652 * DEDUPABLE will be set (NOTE: the UPDATE flag is always set). The caller 1653 * must not modify the data content upon return. 1654 */ 1655 int 1656 hammer2_chain_modify(hammer2_chain_t *chain, hammer2_tid_t mtid, 1657 hammer2_off_t dedup_off, int flags) 1658 { 1659 hammer2_blockref_t obref; 1660 hammer2_dev_t *hmp; 1661 hammer2_io_t *dio; 1662 int error; 1663 int wasinitial; 1664 int setmodified; 1665 int setupdate; 1666 int newmod; 1667 char *bdata; 1668 1669 hmp = chain->hmp; 1670 obref = chain->bref; 1671 KKASSERT(chain->lock.mtx_lock & MTX_EXCLUSIVE); 1672 1673 /* 1674 * Data is not optional for freemap chains (we must always be sure 1675 * to copy the data on COW storage allocations). 1676 */ 1677 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE || 1678 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) { 1679 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) || 1680 (flags & HAMMER2_MODIFY_OPTDATA) == 0); 1681 } 1682 1683 /* 1684 * Data must be resolved if already assigned, unless explicitly 1685 * flagged otherwise. If we cannot safety load the data the 1686 * modification fails and we return early. 1687 */ 1688 if (chain->data == NULL && chain->bytes != 0 && 1689 (flags & HAMMER2_MODIFY_OPTDATA) == 0 && 1690 (chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX)) { 1691 hammer2_chain_load_data(chain); 1692 if (chain->error) 1693 return (chain->error); 1694 } 1695 error = 0; 1696 1697 /* 1698 * Set MODIFIED to indicate that the chain has been modified. A new 1699 * allocation is required when modifying a chain. 1700 * 1701 * Set UPDATE to ensure that the blockref is updated in the parent. 1702 * 1703 * If MODIFIED is already set determine if we can reuse the assigned 1704 * data block or if we need a new data block. 1705 */ 1706 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) { 1707 /* 1708 * Must set modified bit. 1709 */ 1710 atomic_add_long(&hammer2_count_modified_chains, 1); 1711 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1712 hammer2_pfs_memory_inc(chain->pmp); /* can be NULL */ 1713 setmodified = 1; 1714 1715 /* 1716 * We may be able to avoid a copy-on-write if the chain's 1717 * check mode is set to NONE and the chain's current 1718 * modify_tid is beyond the last explicit snapshot tid. 1719 * 1720 * This implements HAMMER2's overwrite-in-place feature. 1721 * 1722 * NOTE! This data-block cannot be used as a de-duplication 1723 * source when the check mode is set to NONE. 1724 */ 1725 if ((chain->bref.type == HAMMER2_BREF_TYPE_DATA || 1726 chain->bref.type == HAMMER2_BREF_TYPE_DIRENT) && 1727 (chain->flags & HAMMER2_CHAIN_INITIAL) == 0 && 1728 (chain->flags & HAMMER2_CHAIN_DEDUPABLE) == 0 && 1729 HAMMER2_DEC_CHECK(chain->bref.methods) == 1730 HAMMER2_CHECK_NONE && 1731 chain->pmp && 1732 chain->bref.modify_tid > 1733 chain->pmp->iroot->meta.pfs_lsnap_tid) { 1734 /* 1735 * Sector overwrite allowed. 1736 */ 1737 newmod = 0; 1738 } else if ((hmp->hflags & HMNT2_EMERG) && 1739 chain->pmp && 1740 chain->bref.modify_tid > 1741 chain->pmp->iroot->meta.pfs_lsnap_tid) { 1742 /* 1743 * If in emergency delete mode then do a modify-in- 1744 * place on any chain type belonging to the PFS as 1745 * long as it doesn't mess up a snapshot. We might 1746 * be forced to do this anyway a little further down 1747 * in the code if the allocation fails. 1748 * 1749 * Also note that in emergency mode, these modify-in- 1750 * place operations are NOT SAFE. A storage failure, 1751 * power failure, or panic can corrupt the filesystem. 1752 */ 1753 newmod = 0; 1754 } else { 1755 /* 1756 * Sector overwrite not allowed, must copy-on-write. 1757 */ 1758 newmod = 1; 1759 } 1760 } else if (chain->flags & HAMMER2_CHAIN_DEDUPABLE) { 1761 /* 1762 * If the modified chain was registered for dedup we need 1763 * a new allocation. This only happens for delayed-flush 1764 * chains (i.e. which run through the front-end buffer 1765 * cache). 1766 */ 1767 newmod = 1; 1768 setmodified = 0; 1769 } else { 1770 /* 1771 * Already flagged modified, no new allocation is needed. 1772 */ 1773 newmod = 0; 1774 setmodified = 0; 1775 } 1776 1777 /* 1778 * Flag parent update required. 1779 */ 1780 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) { 1781 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 1782 setupdate = 1; 1783 } else { 1784 setupdate = 0; 1785 } 1786 1787 /* 1788 * The XOP code returns held but unlocked focus chains. This 1789 * prevents the chain from being destroyed but does not prevent 1790 * it from being modified. diolk is used to interlock modifications 1791 * against XOP frontend accesses to the focus. 1792 * 1793 * This allows us to theoretically avoid deadlocking the frontend 1794 * if one of the backends lock up by not formally locking the 1795 * focused chain in the frontend. In addition, the synchronization 1796 * code relies on this mechanism to avoid deadlocking concurrent 1797 * synchronization threads. 1798 */ 1799 lockmgr(&chain->diolk, LK_EXCLUSIVE); 1800 1801 /* 1802 * The modification or re-modification requires an allocation and 1803 * possible COW. If an error occurs, the previous content and data 1804 * reference is retained and the modification fails. 1805 * 1806 * If dedup_off is non-zero, the caller is requesting a deduplication 1807 * rather than a modification. The MODIFIED bit is not set and the 1808 * data offset is set to the deduplication offset. The data cannot 1809 * be modified. 1810 * 1811 * NOTE: The dedup offset is allowed to be in a partially free state 1812 * and we must be sure to reset it to a fully allocated state 1813 * to force two bulkfree passes to free it again. 1814 * 1815 * NOTE: Only applicable when chain->bytes != 0. 1816 * 1817 * XXX can a chain already be marked MODIFIED without a data 1818 * assignment? If not, assert here instead of testing the case. 1819 */ 1820 if (chain != &hmp->vchain && chain != &hmp->fchain && 1821 chain->bytes) { 1822 if ((chain->bref.data_off & ~HAMMER2_OFF_MASK_RADIX) == 0 || 1823 newmod 1824 ) { 1825 /* 1826 * NOTE: We do not have to remove the dedup 1827 * registration because the area is still 1828 * allocated and the underlying DIO will 1829 * still be flushed. 1830 */ 1831 if (dedup_off) { 1832 chain->bref.data_off = dedup_off; 1833 if ((int)(dedup_off & HAMMER2_OFF_MASK_RADIX)) 1834 chain->bytes = 1 << 1835 (int)(dedup_off & 1836 HAMMER2_OFF_MASK_RADIX); 1837 else 1838 chain->bytes = 0; 1839 chain->error = 0; 1840 atomic_clear_int(&chain->flags, 1841 HAMMER2_CHAIN_MODIFIED); 1842 atomic_add_long(&hammer2_count_modified_chains, 1843 -1); 1844 if (chain->pmp) { 1845 hammer2_pfs_memory_wakeup( 1846 chain->pmp, -1); 1847 } 1848 hammer2_freemap_adjust(hmp, &chain->bref, 1849 HAMMER2_FREEMAP_DORECOVER); 1850 atomic_set_int(&chain->flags, 1851 HAMMER2_CHAIN_DEDUPABLE); 1852 } else { 1853 error = hammer2_freemap_alloc(chain, 1854 chain->bytes); 1855 atomic_clear_int(&chain->flags, 1856 HAMMER2_CHAIN_DEDUPABLE); 1857 1858 /* 1859 * If we are unable to allocate a new block 1860 * but we are in emergency mode, issue a 1861 * warning to the console and reuse the same 1862 * block. 1863 * 1864 * We behave as if the allocation were 1865 * successful. 1866 * 1867 * THIS IS IMPORTANT: These modifications 1868 * are virtually guaranteed to corrupt any 1869 * snapshots related to this filesystem. 1870 */ 1871 if (error && (hmp->hflags & HMNT2_EMERG)) { 1872 error = 0; 1873 chain->bref.flags |= 1874 HAMMER2_BREF_FLAG_EMERG_MIP; 1875 1876 krateprintf(&krate_h2em, 1877 "hammer2: Emergency Mode WARNING: " 1878 "Operation will likely corrupt " 1879 "related snapshot: " 1880 "%016jx.%02x key=%016jx\n", 1881 chain->bref.data_off, 1882 chain->bref.type, 1883 chain->bref.key); 1884 } else if (error == 0) { 1885 chain->bref.flags &= 1886 ~HAMMER2_BREF_FLAG_EMERG_MIP; 1887 } 1888 } 1889 } 1890 } 1891 1892 /* 1893 * Stop here if error. We have to undo any flag bits we might 1894 * have set above. 1895 */ 1896 if (error) { 1897 if (setmodified) { 1898 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 1899 atomic_add_long(&hammer2_count_modified_chains, -1); 1900 if (chain->pmp) 1901 hammer2_pfs_memory_wakeup(chain->pmp, -1); 1902 } 1903 if (setupdate) { 1904 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 1905 } 1906 lockmgr(&chain->diolk, LK_RELEASE); 1907 1908 return error; 1909 } 1910 1911 /* 1912 * Update mirror_tid and modify_tid. modify_tid is only updated 1913 * if not passed as zero (during flushes, parent propagation passes 1914 * the value 0). 1915 * 1916 * NOTE: chain->pmp could be the device spmp. 1917 */ 1918 chain->bref.mirror_tid = hmp->voldata.mirror_tid + 1; 1919 if (mtid) 1920 chain->bref.modify_tid = mtid; 1921 1922 /* 1923 * Set BLKMAPUPD to tell the flush code that an existing blockmap entry 1924 * requires updating as well as to tell the delete code that the 1925 * chain's blockref might not exactly match (in terms of physical size 1926 * or block offset) the one in the parent's blocktable. The base key 1927 * of course will still match. 1928 */ 1929 if (chain->flags & HAMMER2_CHAIN_BLKMAPPED) 1930 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPUPD); 1931 1932 /* 1933 * Short-cut data block handling when the caller does not need an 1934 * actual data reference to (aka OPTDATA), as long as the chain does 1935 * not already have a data pointer to the data and no de-duplication 1936 * occurred. 1937 * 1938 * This generally means that the modifications are being done via the 1939 * logical buffer cache. 1940 * 1941 * NOTE: If deduplication occurred we have to run through the data 1942 * stuff to clear INITIAL, and the caller will likely want to 1943 * assign the check code anyway. Leaving INITIAL set on a 1944 * dedup can be deadly (it can cause the block to be zero'd!). 1945 * 1946 * This code also handles bytes == 0 (most dirents). 1947 */ 1948 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA && 1949 (flags & HAMMER2_MODIFY_OPTDATA) && 1950 chain->data == NULL) { 1951 if (dedup_off == 0) { 1952 KKASSERT(chain->dio == NULL); 1953 goto skip2; 1954 } 1955 } 1956 1957 /* 1958 * Clearing the INITIAL flag (for indirect blocks) indicates that 1959 * we've processed the uninitialized storage allocation. 1960 * 1961 * If this flag is already clear we are likely in a copy-on-write 1962 * situation but we have to be sure NOT to bzero the storage if 1963 * no data is present. 1964 * 1965 * Clearing of NOTTESTED is allowed if the MODIFIED bit is set, 1966 */ 1967 if (chain->flags & HAMMER2_CHAIN_INITIAL) { 1968 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1969 wasinitial = 1; 1970 } else { 1971 wasinitial = 0; 1972 } 1973 1974 /* 1975 * Instantiate data buffer and possibly execute COW operation 1976 */ 1977 switch(chain->bref.type) { 1978 case HAMMER2_BREF_TYPE_VOLUME: 1979 case HAMMER2_BREF_TYPE_FREEMAP: 1980 /* 1981 * The data is embedded, no copy-on-write operation is 1982 * needed. 1983 */ 1984 KKASSERT(chain->dio == NULL); 1985 break; 1986 case HAMMER2_BREF_TYPE_DIRENT: 1987 /* 1988 * The data might be fully embedded. 1989 */ 1990 if (chain->bytes == 0) { 1991 KKASSERT(chain->dio == NULL); 1992 break; 1993 } 1994 /* fall through */ 1995 case HAMMER2_BREF_TYPE_INODE: 1996 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1997 case HAMMER2_BREF_TYPE_DATA: 1998 case HAMMER2_BREF_TYPE_INDIRECT: 1999 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2000 /* 2001 * Perform the copy-on-write operation 2002 * 2003 * zero-fill or copy-on-write depending on whether 2004 * chain->data exists or not and set the dirty state for 2005 * the new buffer. hammer2_io_new() will handle the 2006 * zero-fill. 2007 * 2008 * If a dedup_off was supplied this is an existing block 2009 * and no COW, copy, or further modification is required. 2010 */ 2011 KKASSERT(chain != &hmp->vchain && chain != &hmp->fchain); 2012 2013 if (wasinitial && dedup_off == 0) { 2014 error = hammer2_io_new(hmp, chain->bref.type, 2015 chain->bref.data_off, 2016 chain->bytes, &dio); 2017 } else { 2018 error = hammer2_io_bread(hmp, chain->bref.type, 2019 chain->bref.data_off, 2020 chain->bytes, &dio); 2021 } 2022 hammer2_adjreadcounter(chain->bref.type, chain->bytes); 2023 2024 /* 2025 * If an I/O error occurs make sure callers cannot accidently 2026 * modify the old buffer's contents and corrupt the filesystem. 2027 * 2028 * NOTE: hammer2_io_data() call issues bkvasync() 2029 */ 2030 if (error) { 2031 kprintf("hammer2_chain_modify: hmp=%p I/O error\n", 2032 hmp); 2033 chain->error = HAMMER2_ERROR_EIO; 2034 hammer2_io_brelse(&dio); 2035 hammer2_io_brelse(&chain->dio); 2036 chain->data = NULL; 2037 break; 2038 } 2039 chain->error = 0; 2040 bdata = hammer2_io_data(dio, chain->bref.data_off); 2041 2042 if (chain->data) { 2043 /* 2044 * COW (unless a dedup). 2045 */ 2046 KKASSERT(chain->dio != NULL); 2047 if (chain->data != (void *)bdata && dedup_off == 0) { 2048 bcopy(chain->data, bdata, chain->bytes); 2049 } 2050 } else if (wasinitial == 0 && dedup_off == 0) { 2051 /* 2052 * We have a problem. We were asked to COW but 2053 * we don't have any data to COW with! 2054 */ 2055 panic("hammer2_chain_modify: having a COW %p\n", 2056 chain); 2057 } 2058 2059 /* 2060 * Retire the old buffer, replace with the new. Dirty or 2061 * redirty the new buffer. 2062 * 2063 * WARNING! The system buffer cache may have already flushed 2064 * the buffer, so we must be sure to [re]dirty it 2065 * for further modification. 2066 * 2067 * If dedup_off was supplied, the caller is not 2068 * expected to make any further modification to the 2069 * buffer. 2070 * 2071 * WARNING! hammer2_get_gdata() assumes dio never transitions 2072 * through NULL in order to optimize away unnecessary 2073 * diolk operations. 2074 */ 2075 { 2076 hammer2_io_t *tio; 2077 2078 if ((tio = chain->dio) != NULL) 2079 hammer2_io_bqrelse(&tio); 2080 chain->data = (void *)bdata; 2081 chain->dio = dio; 2082 if (dedup_off == 0) 2083 hammer2_io_setdirty(dio); 2084 } 2085 break; 2086 default: 2087 panic("hammer2_chain_modify: illegal non-embedded type %d", 2088 chain->bref.type); 2089 break; 2090 2091 } 2092 skip2: 2093 /* 2094 * setflush on parent indicating that the parent must recurse down 2095 * to us. Do not call on chain itself which might already have it 2096 * set. 2097 */ 2098 if (chain->parent) 2099 hammer2_chain_setflush(chain->parent); 2100 lockmgr(&chain->diolk, LK_RELEASE); 2101 2102 return (chain->error); 2103 } 2104 2105 /* 2106 * Modify the chain associated with an inode. 2107 */ 2108 int 2109 hammer2_chain_modify_ip(hammer2_inode_t *ip, hammer2_chain_t *chain, 2110 hammer2_tid_t mtid, int flags) 2111 { 2112 int error; 2113 2114 hammer2_inode_modify(ip); 2115 error = hammer2_chain_modify(chain, mtid, 0, flags); 2116 2117 return error; 2118 } 2119 2120 /* 2121 * This function returns the chain at the nearest key within the specified 2122 * range. The returned chain will be referenced but not locked. 2123 * 2124 * This function will recurse through chain->rbtree as necessary and will 2125 * return a *key_nextp suitable for iteration. *key_nextp is only set if 2126 * the iteration value is less than the current value of *key_nextp. 2127 * 2128 * The caller should use (*key_nextp) to calculate the actual range of 2129 * the returned element, which will be (key_beg to *key_nextp - 1), because 2130 * there might be another element which is superior to the returned element 2131 * and overlaps it. 2132 * 2133 * (*key_nextp) can be passed as key_beg in an iteration only while non-NULL 2134 * chains continue to be returned. On EOF (*key_nextp) may overflow since 2135 * it will wind up being (key_end + 1). 2136 * 2137 * WARNING! Must be called with child's spinlock held. Spinlock remains 2138 * held through the operation. 2139 */ 2140 struct hammer2_chain_find_info { 2141 hammer2_chain_t *best; 2142 hammer2_key_t key_beg; 2143 hammer2_key_t key_end; 2144 hammer2_key_t key_next; 2145 }; 2146 2147 static int hammer2_chain_find_cmp(hammer2_chain_t *child, void *data); 2148 static int hammer2_chain_find_callback(hammer2_chain_t *child, void *data); 2149 2150 static 2151 hammer2_chain_t * 2152 hammer2_chain_find(hammer2_chain_t *parent, hammer2_key_t *key_nextp, 2153 hammer2_key_t key_beg, hammer2_key_t key_end) 2154 { 2155 struct hammer2_chain_find_info info; 2156 2157 info.best = NULL; 2158 info.key_beg = key_beg; 2159 info.key_end = key_end; 2160 info.key_next = *key_nextp; 2161 2162 RB_SCAN(hammer2_chain_tree, &parent->core.rbtree, 2163 hammer2_chain_find_cmp, hammer2_chain_find_callback, 2164 &info); 2165 *key_nextp = info.key_next; 2166 #if 0 2167 kprintf("chain_find %p %016jx:%016jx next=%016jx\n", 2168 parent, key_beg, key_end, *key_nextp); 2169 #endif 2170 2171 return (info.best); 2172 } 2173 2174 static 2175 int 2176 hammer2_chain_find_cmp(hammer2_chain_t *child, void *data) 2177 { 2178 struct hammer2_chain_find_info *info = data; 2179 hammer2_key_t child_beg; 2180 hammer2_key_t child_end; 2181 2182 child_beg = child->bref.key; 2183 child_end = child_beg + ((hammer2_key_t)1 << child->bref.keybits) - 1; 2184 2185 if (child_end < info->key_beg) 2186 return(-1); 2187 if (child_beg > info->key_end) 2188 return(1); 2189 return(0); 2190 } 2191 2192 static 2193 int 2194 hammer2_chain_find_callback(hammer2_chain_t *child, void *data) 2195 { 2196 struct hammer2_chain_find_info *info = data; 2197 hammer2_chain_t *best; 2198 hammer2_key_t child_end; 2199 2200 /* 2201 * WARNING! Layerq is scanned forwards, exact matches should keep 2202 * the existing info->best. 2203 */ 2204 if ((best = info->best) == NULL) { 2205 /* 2206 * No previous best. Assign best 2207 */ 2208 info->best = child; 2209 } else if (best->bref.key <= info->key_beg && 2210 child->bref.key <= info->key_beg) { 2211 /* 2212 * Illegal overlap. 2213 */ 2214 KKASSERT(0); 2215 /*info->best = child;*/ 2216 } else if (child->bref.key < best->bref.key) { 2217 /* 2218 * Child has a nearer key and best is not flush with key_beg. 2219 * Set best to child. Truncate key_next to the old best key. 2220 */ 2221 info->best = child; 2222 if (info->key_next > best->bref.key || info->key_next == 0) 2223 info->key_next = best->bref.key; 2224 } else if (child->bref.key == best->bref.key) { 2225 /* 2226 * If our current best is flush with the child then this 2227 * is an illegal overlap. 2228 * 2229 * key_next will automatically be limited to the smaller of 2230 * the two end-points. 2231 */ 2232 KKASSERT(0); 2233 info->best = child; 2234 } else { 2235 /* 2236 * Keep the current best but truncate key_next to the child's 2237 * base. 2238 * 2239 * key_next will also automatically be limited to the smaller 2240 * of the two end-points (probably not necessary for this case 2241 * but we do it anyway). 2242 */ 2243 if (info->key_next > child->bref.key || info->key_next == 0) 2244 info->key_next = child->bref.key; 2245 } 2246 2247 /* 2248 * Always truncate key_next based on child's end-of-range. 2249 */ 2250 child_end = child->bref.key + ((hammer2_key_t)1 << child->bref.keybits); 2251 if (child_end && (info->key_next > child_end || info->key_next == 0)) 2252 info->key_next = child_end; 2253 2254 return(0); 2255 } 2256 2257 /* 2258 * Retrieve the specified chain from a media blockref, creating the 2259 * in-memory chain structure which reflects it. The returned chain is 2260 * held and locked according to (how) (HAMMER2_RESOLVE_*). The caller must 2261 * handle crc-checks and so forth, and should check chain->error before 2262 * assuming that the data is good. 2263 * 2264 * To handle insertion races pass the INSERT_RACE flag along with the 2265 * generation number of the core. NULL will be returned if the generation 2266 * number changes before we have a chance to insert the chain. Insert 2267 * races can occur because the parent might be held shared. 2268 * 2269 * Caller must hold the parent locked shared or exclusive since we may 2270 * need the parent's bref array to find our block. 2271 * 2272 * WARNING! chain->pmp is always set to NULL for any chain representing 2273 * part of the super-root topology. 2274 */ 2275 hammer2_chain_t * 2276 hammer2_chain_get(hammer2_chain_t *parent, int generation, 2277 hammer2_blockref_t *bref, int how) 2278 { 2279 hammer2_dev_t *hmp = parent->hmp; 2280 hammer2_chain_t *chain; 2281 int error; 2282 2283 /* 2284 * Allocate a chain structure representing the existing media 2285 * entry. Resulting chain has one ref and is not locked. 2286 */ 2287 if (bref->flags & HAMMER2_BREF_FLAG_PFSROOT) 2288 chain = hammer2_chain_alloc(hmp, NULL, bref); 2289 else 2290 chain = hammer2_chain_alloc(hmp, parent->pmp, bref); 2291 /* ref'd chain returned */ 2292 2293 /* 2294 * Flag that the chain is in the parent's blockmap so delete/flush 2295 * knows what to do with it. 2296 */ 2297 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED); 2298 2299 /* 2300 * chain must be locked to avoid unexpected ripouts 2301 */ 2302 hammer2_chain_lock(chain, how); 2303 2304 /* 2305 * Link the chain into its parent. A spinlock is required to safely 2306 * access the RBTREE, and it is possible to collide with another 2307 * hammer2_chain_get() operation because the caller might only hold 2308 * a shared lock on the parent. 2309 * 2310 * NOTE: Get races can occur quite often when we distribute 2311 * asynchronous read-aheads across multiple threads. 2312 */ 2313 KKASSERT(parent->refs > 0); 2314 error = hammer2_chain_insert(parent, chain, 2315 HAMMER2_CHAIN_INSERT_SPIN | 2316 HAMMER2_CHAIN_INSERT_RACE, 2317 generation); 2318 if (error) { 2319 KKASSERT((chain->flags & HAMMER2_CHAIN_ONRBTREE) == 0); 2320 /*kprintf("chain %p get race\n", chain);*/ 2321 hammer2_chain_unlock(chain); 2322 hammer2_chain_drop(chain); 2323 chain = NULL; 2324 } else { 2325 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE); 2326 } 2327 2328 /* 2329 * Return our new chain referenced but not locked, or NULL if 2330 * a race occurred. 2331 */ 2332 return (chain); 2333 } 2334 2335 /* 2336 * Lookup initialization/completion API 2337 */ 2338 hammer2_chain_t * 2339 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags) 2340 { 2341 hammer2_chain_ref(parent); 2342 if (flags & HAMMER2_LOOKUP_SHARED) { 2343 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS | 2344 HAMMER2_RESOLVE_SHARED); 2345 } else { 2346 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 2347 } 2348 return (parent); 2349 } 2350 2351 void 2352 hammer2_chain_lookup_done(hammer2_chain_t *parent) 2353 { 2354 if (parent) { 2355 hammer2_chain_unlock(parent); 2356 hammer2_chain_drop(parent); 2357 } 2358 } 2359 2360 /* 2361 * Take the locked chain and return a locked parent. The chain remains 2362 * locked on return, but may have to be temporarily unlocked to acquire 2363 * the parent. Because of this, (chain) must be stable and cannot be 2364 * deleted while it was temporarily unlocked (typically means that (chain) 2365 * is an inode). 2366 * 2367 * Pass HAMMER2_RESOLVE_* flags in flags. 2368 * 2369 * This will work even if the chain is errored, and the caller can check 2370 * parent->error on return if desired since the parent will be locked. 2371 * 2372 * This function handles the lock order reversal. 2373 */ 2374 hammer2_chain_t * 2375 hammer2_chain_getparent(hammer2_chain_t *chain, int flags) 2376 { 2377 hammer2_chain_t *parent; 2378 2379 /* 2380 * Be careful of order, chain must be unlocked before parent 2381 * is locked below to avoid a deadlock. Try it trivially first. 2382 */ 2383 parent = chain->parent; 2384 if (parent == NULL) 2385 panic("hammer2_chain_getparent: no parent"); 2386 hammer2_chain_ref(parent); 2387 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) 2388 return parent; 2389 2390 for (;;) { 2391 hammer2_chain_unlock(chain); 2392 hammer2_chain_lock(parent, flags); 2393 hammer2_chain_lock(chain, flags); 2394 2395 /* 2396 * Parent relinking races are quite common. We have to get 2397 * it right or we will blow up the block table. 2398 */ 2399 if (chain->parent == parent) 2400 break; 2401 hammer2_chain_unlock(parent); 2402 hammer2_chain_drop(parent); 2403 cpu_ccfence(); 2404 parent = chain->parent; 2405 if (parent == NULL) 2406 panic("hammer2_chain_getparent: no parent"); 2407 hammer2_chain_ref(parent); 2408 } 2409 return parent; 2410 } 2411 2412 /* 2413 * Take the locked chain and return a locked parent. The chain is unlocked 2414 * and dropped. *chainp is set to the returned parent as a convenience. 2415 * Pass HAMMER2_RESOLVE_* flags in flags. 2416 * 2417 * This will work even if the chain is errored, and the caller can check 2418 * parent->error on return if desired since the parent will be locked. 2419 * 2420 * The chain does NOT need to be stable. We use a tracking structure 2421 * to track the expected parent if the chain is deleted out from under us. 2422 * 2423 * This function handles the lock order reversal. 2424 */ 2425 hammer2_chain_t * 2426 hammer2_chain_repparent(hammer2_chain_t **chainp, int flags) 2427 { 2428 hammer2_chain_t *chain; 2429 hammer2_chain_t *parent; 2430 struct hammer2_reptrack reptrack; 2431 struct hammer2_reptrack **repp; 2432 2433 /* 2434 * Be careful of order, chain must be unlocked before parent 2435 * is locked below to avoid a deadlock. Try it trivially first. 2436 */ 2437 chain = *chainp; 2438 parent = chain->parent; 2439 if (parent == NULL) { 2440 hammer2_spin_unex(&chain->core.spin); 2441 panic("hammer2_chain_repparent: no parent"); 2442 } 2443 hammer2_chain_ref(parent); 2444 if (hammer2_chain_lock(parent, flags|HAMMER2_RESOLVE_NONBLOCK) == 0) { 2445 hammer2_chain_unlock(chain); 2446 hammer2_chain_drop(chain); 2447 *chainp = parent; 2448 2449 return parent; 2450 } 2451 2452 /* 2453 * Ok, now it gets a bit nasty. There are multiple situations where 2454 * the parent might be in the middle of a deletion, or where the child 2455 * (chain) might be deleted the instant we let go of its lock. 2456 * We can potentially end up in a no-win situation! 2457 * 2458 * In particular, the indirect_maintenance() case can cause these 2459 * situations. 2460 * 2461 * To deal with this we install a reptrack structure in the parent 2462 * This reptrack structure 'owns' the parent ref and will automatically 2463 * migrate to the parent's parent if the parent is deleted permanently. 2464 */ 2465 hammer2_spin_init(&reptrack.spin, "h2reptrk"); 2466 reptrack.chain = parent; 2467 hammer2_chain_ref(parent); /* for the reptrack */ 2468 2469 hammer2_spin_ex(&parent->core.spin); 2470 reptrack.next = parent->core.reptrack; 2471 parent->core.reptrack = &reptrack; 2472 hammer2_spin_unex(&parent->core.spin); 2473 2474 hammer2_chain_unlock(chain); 2475 hammer2_chain_drop(chain); 2476 chain = NULL; /* gone */ 2477 2478 /* 2479 * At the top of this loop, chain is gone and parent is refd both 2480 * by us explicitly AND via our reptrack. We are attempting to 2481 * lock parent. 2482 */ 2483 for (;;) { 2484 hammer2_chain_lock(parent, flags); 2485 2486 if (reptrack.chain == parent) 2487 break; 2488 hammer2_chain_unlock(parent); 2489 hammer2_chain_drop(parent); 2490 2491 kprintf("hammer2: debug REPTRACK %p->%p\n", 2492 parent, reptrack.chain); 2493 hammer2_spin_ex(&reptrack.spin); 2494 parent = reptrack.chain; 2495 hammer2_chain_ref(parent); 2496 hammer2_spin_unex(&reptrack.spin); 2497 } 2498 2499 /* 2500 * Once parent is locked and matches our reptrack, our reptrack 2501 * will be stable and we have our parent. We can unlink our 2502 * reptrack. 2503 * 2504 * WARNING! Remember that the chain lock might be shared. Chains 2505 * locked shared have stable parent linkages. 2506 */ 2507 hammer2_spin_ex(&parent->core.spin); 2508 repp = &parent->core.reptrack; 2509 while (*repp != &reptrack) 2510 repp = &(*repp)->next; 2511 *repp = reptrack.next; 2512 hammer2_spin_unex(&parent->core.spin); 2513 2514 hammer2_chain_drop(parent); /* reptrack ref */ 2515 *chainp = parent; /* return parent lock+ref */ 2516 2517 return parent; 2518 } 2519 2520 /* 2521 * Dispose of any linked reptrack structures in (chain) by shifting them to 2522 * (parent). Both (chain) and (parent) must be exclusively locked. 2523 * 2524 * This is interlocked against any children of (chain) on the other side. 2525 * No children so remain as-of when this is called so we can test 2526 * core.reptrack without holding the spin-lock. 2527 * 2528 * Used whenever the caller intends to permanently delete chains related 2529 * to topological recursions (BREF_TYPE_INDIRECT, BREF_TYPE_FREEMAP_NODE), 2530 * where the chains underneath the node being deleted are given a new parent 2531 * above the node being deleted. 2532 */ 2533 static 2534 void 2535 hammer2_chain_repchange(hammer2_chain_t *parent, hammer2_chain_t *chain) 2536 { 2537 struct hammer2_reptrack *reptrack; 2538 2539 KKASSERT(chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)); 2540 while (chain->core.reptrack) { 2541 hammer2_spin_ex(&parent->core.spin); 2542 hammer2_spin_ex(&chain->core.spin); 2543 reptrack = chain->core.reptrack; 2544 if (reptrack == NULL) { 2545 hammer2_spin_unex(&chain->core.spin); 2546 hammer2_spin_unex(&parent->core.spin); 2547 break; 2548 } 2549 hammer2_spin_ex(&reptrack->spin); 2550 chain->core.reptrack = reptrack->next; 2551 reptrack->chain = parent; 2552 reptrack->next = parent->core.reptrack; 2553 parent->core.reptrack = reptrack; 2554 hammer2_chain_ref(parent); /* reptrack */ 2555 2556 hammer2_spin_unex(&chain->core.spin); 2557 hammer2_spin_unex(&parent->core.spin); 2558 kprintf("hammer2: debug repchange %p %p->%p\n", 2559 reptrack, chain, parent); 2560 hammer2_chain_drop(chain); /* reptrack */ 2561 } 2562 } 2563 2564 /* 2565 * Locate the first chain whos key range overlaps (key_beg, key_end) inclusive. 2566 * (*parentp) typically points to an inode but can also point to a related 2567 * indirect block and this function will recurse upwards and find the inode 2568 * or the nearest undeleted indirect block covering the key range. 2569 * 2570 * This function unconditionally sets *errorp, replacing any previous value. 2571 * 2572 * (*parentp) must be exclusive or shared locked (depending on flags) and 2573 * referenced and can be an inode or an existing indirect block within the 2574 * inode. 2575 * 2576 * If (*parent) is errored out, this function will not attempt to recurse 2577 * the radix tree and will return NULL along with an appropriate *errorp. 2578 * If NULL is returned and *errorp is 0, the requested lookup could not be 2579 * located. 2580 * 2581 * On return (*parentp) will be modified to point at the deepest parent chain 2582 * element encountered during the search, as a helper for an insertion or 2583 * deletion. 2584 * 2585 * The new (*parentp) will be locked shared or exclusive (depending on flags), 2586 * and referenced, and the old will be unlocked and dereferenced (no change 2587 * if they are both the same). This is particularly important if the caller 2588 * wishes to insert a new chain, (*parentp) will be set properly even if NULL 2589 * is returned, as long as no error occurred. 2590 * 2591 * The matching chain will be returned locked according to flags. 2592 * 2593 * -- 2594 * 2595 * NULL is returned if no match was found, but (*parentp) will still 2596 * potentially be adjusted. 2597 * 2598 * On return (*key_nextp) will point to an iterative value for key_beg. 2599 * (If NULL is returned (*key_nextp) is set to (key_end + 1)). 2600 * 2601 * This function will also recurse up the chain if the key is not within the 2602 * current parent's range. (*parentp) can never be set to NULL. An iteration 2603 * can simply allow (*parentp) to float inside the loop. 2604 * 2605 * NOTE! chain->data is not always resolved. By default it will not be 2606 * resolved for BREF_TYPE_DATA, FREEMAP_NODE, or FREEMAP_LEAF. Use 2607 * HAMMER2_LOOKUP_ALWAYS to force resolution (but be careful w/ 2608 * BREF_TYPE_DATA as the device buffer can alias the logical file 2609 * buffer). 2610 */ 2611 hammer2_chain_t * 2612 hammer2_chain_lookup(hammer2_chain_t **parentp, hammer2_key_t *key_nextp, 2613 hammer2_key_t key_beg, hammer2_key_t key_end, 2614 int *errorp, int flags) 2615 { 2616 hammer2_dev_t *hmp; 2617 hammer2_chain_t *parent; 2618 hammer2_chain_t *chain; 2619 hammer2_blockref_t *base; 2620 hammer2_blockref_t *bref; 2621 hammer2_blockref_t bsave; 2622 hammer2_key_t scan_beg; 2623 hammer2_key_t scan_end; 2624 int count = 0; 2625 int how_always = HAMMER2_RESOLVE_ALWAYS; 2626 int how_maybe = HAMMER2_RESOLVE_MAYBE; 2627 int how; 2628 int generation; 2629 int maxloops = 300000; 2630 volatile hammer2_mtx_t save_mtx; 2631 2632 if (flags & HAMMER2_LOOKUP_ALWAYS) { 2633 how_maybe = how_always; 2634 how = HAMMER2_RESOLVE_ALWAYS; 2635 } else if (flags & HAMMER2_LOOKUP_NODATA) { 2636 how = HAMMER2_RESOLVE_NEVER; 2637 } else { 2638 how = HAMMER2_RESOLVE_MAYBE; 2639 } 2640 if (flags & HAMMER2_LOOKUP_SHARED) { 2641 how_maybe |= HAMMER2_RESOLVE_SHARED; 2642 how_always |= HAMMER2_RESOLVE_SHARED; 2643 how |= HAMMER2_RESOLVE_SHARED; 2644 } 2645 2646 /* 2647 * Recurse (*parentp) upward if necessary until the parent completely 2648 * encloses the key range or we hit the inode. 2649 * 2650 * Handle races against the flusher deleting indirect nodes on its 2651 * way back up by continuing to recurse upward past the deletion. 2652 */ 2653 parent = *parentp; 2654 hmp = parent->hmp; 2655 *errorp = 0; 2656 2657 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 2658 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2659 scan_beg = parent->bref.key; 2660 scan_end = scan_beg + 2661 ((hammer2_key_t)1 << parent->bref.keybits) - 1; 2662 if ((parent->flags & HAMMER2_CHAIN_DELETED) == 0) { 2663 if (key_beg >= scan_beg && key_end <= scan_end) 2664 break; 2665 } 2666 parent = hammer2_chain_repparent(parentp, how_maybe); 2667 } 2668 again: 2669 if (--maxloops == 0) 2670 panic("hammer2_chain_lookup: maxloops"); 2671 2672 /* 2673 * MATCHIND case that does not require parent->data (do prior to 2674 * parent->error check). 2675 */ 2676 switch(parent->bref.type) { 2677 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2678 case HAMMER2_BREF_TYPE_INDIRECT: 2679 if (flags & HAMMER2_LOOKUP_MATCHIND) { 2680 scan_beg = parent->bref.key; 2681 scan_end = scan_beg + 2682 ((hammer2_key_t)1 << parent->bref.keybits) - 1; 2683 if (key_beg == scan_beg && key_end == scan_end) { 2684 chain = parent; 2685 hammer2_chain_ref(chain); 2686 hammer2_chain_lock(chain, how_maybe); 2687 *key_nextp = scan_end + 1; 2688 goto done; 2689 } 2690 } 2691 break; 2692 default: 2693 break; 2694 } 2695 2696 /* 2697 * No lookup is possible if the parent is errored. We delayed 2698 * this check as long as we could to ensure that the parent backup, 2699 * embedded data, and MATCHIND code could still execute. 2700 */ 2701 if (parent->error) { 2702 *errorp = parent->error; 2703 return NULL; 2704 } 2705 2706 /* 2707 * Locate the blockref array. Currently we do a fully associative 2708 * search through the array. 2709 */ 2710 switch(parent->bref.type) { 2711 case HAMMER2_BREF_TYPE_INODE: 2712 /* 2713 * Special shortcut for embedded data returns the inode 2714 * itself. Callers must detect this condition and access 2715 * the embedded data (the strategy code does this for us). 2716 * 2717 * This is only applicable to regular files and softlinks. 2718 * 2719 * We need a second lock on parent. Since we already have 2720 * a lock we must pass LOCKAGAIN to prevent unexpected 2721 * blocking (we don't want to block on a second shared 2722 * ref if an exclusive lock is pending) 2723 */ 2724 if (parent->data->ipdata.meta.op_flags & 2725 HAMMER2_OPFLAG_DIRECTDATA) { 2726 if (flags & HAMMER2_LOOKUP_NODIRECT) { 2727 chain = NULL; 2728 *key_nextp = key_end + 1; 2729 goto done; 2730 } 2731 hammer2_chain_ref(parent); 2732 hammer2_chain_lock(parent, how_always | 2733 HAMMER2_RESOLVE_LOCKAGAIN); 2734 *key_nextp = key_end + 1; 2735 return (parent); 2736 } 2737 base = &parent->data->ipdata.u.blockset.blockref[0]; 2738 count = HAMMER2_SET_COUNT; 2739 break; 2740 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2741 case HAMMER2_BREF_TYPE_INDIRECT: 2742 /* 2743 * Optimize indirect blocks in the INITIAL state to avoid 2744 * I/O. 2745 * 2746 * Debugging: Enter permanent wait state instead of 2747 * panicing on unexpectedly NULL data for the moment. 2748 */ 2749 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 2750 base = NULL; 2751 } else { 2752 if (parent->data == NULL) { 2753 kprintf("hammer2: unexpected NULL data " 2754 "on %p\n", parent); 2755 while (1) 2756 tsleep(parent, 0, "xxx", 0); 2757 } 2758 base = &parent->data->npdata[0]; 2759 } 2760 count = parent->bytes / sizeof(hammer2_blockref_t); 2761 break; 2762 case HAMMER2_BREF_TYPE_VOLUME: 2763 base = &parent->data->voldata.sroot_blockset.blockref[0]; 2764 count = HAMMER2_SET_COUNT; 2765 break; 2766 case HAMMER2_BREF_TYPE_FREEMAP: 2767 base = &parent->data->blkset.blockref[0]; 2768 count = HAMMER2_SET_COUNT; 2769 break; 2770 default: 2771 panic("hammer2_chain_lookup: unrecognized " 2772 "blockref(B) type: %d", 2773 parent->bref.type); 2774 base = NULL; /* safety */ 2775 count = 0; /* safety */ 2776 break; 2777 } 2778 2779 /* 2780 * Merged scan to find next candidate. 2781 * 2782 * hammer2_base_*() functions require the parent->core.live_* fields 2783 * to be synchronized. 2784 * 2785 * We need to hold the spinlock to access the block array and RB tree 2786 * and to interlock chain creation. 2787 */ 2788 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) 2789 hammer2_chain_countbrefs(parent, base, count); 2790 2791 /* 2792 * Combined search 2793 */ 2794 hammer2_spin_ex(&parent->core.spin); 2795 chain = hammer2_combined_find(parent, base, count, 2796 key_nextp, 2797 key_beg, key_end, 2798 &bref); 2799 generation = parent->core.generation; 2800 2801 /* 2802 * Exhausted parent chain, iterate. 2803 */ 2804 if (bref == NULL) { 2805 KKASSERT(chain == NULL); 2806 hammer2_spin_unex(&parent->core.spin); 2807 if (key_beg == key_end) /* short cut single-key case */ 2808 return (NULL); 2809 2810 /* 2811 * Stop if we reached the end of the iteration. 2812 */ 2813 if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT && 2814 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2815 return (NULL); 2816 } 2817 2818 /* 2819 * Calculate next key, stop if we reached the end of the 2820 * iteration, otherwise go up one level and loop. 2821 */ 2822 key_beg = parent->bref.key + 2823 ((hammer2_key_t)1 << parent->bref.keybits); 2824 if (key_beg == 0 || key_beg > key_end) 2825 return (NULL); 2826 parent = hammer2_chain_repparent(parentp, how_maybe); 2827 goto again; 2828 } 2829 2830 /* 2831 * Selected from blockref or in-memory chain. 2832 */ 2833 bsave = *bref; 2834 if (chain == NULL) { 2835 hammer2_spin_unex(&parent->core.spin); 2836 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT || 2837 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2838 chain = hammer2_chain_get(parent, generation, 2839 &bsave, how_maybe); 2840 } else { 2841 chain = hammer2_chain_get(parent, generation, 2842 &bsave, how); 2843 } 2844 if (chain == NULL) 2845 goto again; 2846 } else { 2847 hammer2_chain_ref(chain); 2848 hammer2_spin_unex(&parent->core.spin); 2849 2850 /* 2851 * chain is referenced but not locked. We must lock the 2852 * chain to obtain definitive state. 2853 */ 2854 if (bsave.type == HAMMER2_BREF_TYPE_INDIRECT || 2855 bsave.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2856 hammer2_chain_lock(chain, how_maybe); 2857 } else { 2858 hammer2_chain_lock(chain, how); 2859 } 2860 KKASSERT(chain->parent == parent); 2861 } 2862 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) || 2863 chain->parent != parent) { 2864 hammer2_chain_unlock(chain); 2865 hammer2_chain_drop(chain); 2866 chain = NULL; /* SAFETY */ 2867 goto again; 2868 } 2869 2870 2871 /* 2872 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX) 2873 * 2874 * NOTE: Chain's key range is not relevant as there might be 2875 * one-offs within the range that are not deleted. 2876 * 2877 * NOTE: Lookups can race delete-duplicate because 2878 * delete-duplicate does not lock the parent's core 2879 * (they just use the spinlock on the core). 2880 */ 2881 if (chain->flags & HAMMER2_CHAIN_DELETED) { 2882 kprintf("skip deleted chain %016jx.%02x key=%016jx\n", 2883 chain->bref.data_off, chain->bref.type, 2884 chain->bref.key); 2885 hammer2_chain_unlock(chain); 2886 hammer2_chain_drop(chain); 2887 chain = NULL; /* SAFETY */ 2888 key_beg = *key_nextp; 2889 if (key_beg == 0 || key_beg > key_end) 2890 return(NULL); 2891 goto again; 2892 } 2893 2894 /* 2895 * If the chain element is an indirect block it becomes the new 2896 * parent and we loop on it. We must maintain our top-down locks 2897 * to prevent the flusher from interfering (i.e. doing a 2898 * delete-duplicate and leaving us recursing down a deleted chain). 2899 * 2900 * The parent always has to be locked with at least RESOLVE_MAYBE 2901 * so we can access its data. It might need a fixup if the caller 2902 * passed incompatible flags. Be careful not to cause a deadlock 2903 * as a data-load requires an exclusive lock. 2904 * 2905 * If HAMMER2_LOOKUP_MATCHIND is set and the indirect block's key 2906 * range is within the requested key range we return the indirect 2907 * block and do NOT loop. This is usually only used to acquire 2908 * freemap nodes. 2909 */ 2910 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 2911 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2912 save_mtx = parent->lock; 2913 hammer2_chain_unlock(parent); 2914 hammer2_chain_drop(parent); 2915 *parentp = parent = chain; 2916 chain = NULL; /* SAFETY */ 2917 goto again; 2918 } 2919 done: 2920 /* 2921 * All done, return the locked chain. 2922 * 2923 * If the caller does not want a locked chain, replace the lock with 2924 * a ref. Perhaps this can eventually be optimized to not obtain the 2925 * lock in the first place for situations where the data does not 2926 * need to be resolved. 2927 * 2928 * NOTE! A chain->error must be tested by the caller upon return. 2929 * *errorp is only set based on issues which occur while 2930 * trying to reach the chain. 2931 */ 2932 return (chain); 2933 } 2934 2935 /* 2936 * After having issued a lookup we can iterate all matching keys. 2937 * 2938 * If chain is non-NULL we continue the iteration from just after it's index. 2939 * 2940 * If chain is NULL we assume the parent was exhausted and continue the 2941 * iteration at the next parent. 2942 * 2943 * If a fatal error occurs (typically an I/O error), a dummy chain is 2944 * returned with chain->error and error-identifying information set. This 2945 * chain will assert if you try to do anything fancy with it. 2946 * 2947 * XXX Depending on where the error occurs we should allow continued iteration. 2948 * 2949 * parent must be locked on entry and remains locked throughout. chain's 2950 * lock status must match flags. Chain is always at least referenced. 2951 * 2952 * WARNING! The MATCHIND flag does not apply to this function. 2953 */ 2954 hammer2_chain_t * 2955 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain, 2956 hammer2_key_t *key_nextp, 2957 hammer2_key_t key_beg, hammer2_key_t key_end, 2958 int *errorp, int flags) 2959 { 2960 hammer2_chain_t *parent; 2961 int how_maybe; 2962 2963 /* 2964 * Calculate locking flags for upward recursion. 2965 */ 2966 how_maybe = HAMMER2_RESOLVE_MAYBE; 2967 if (flags & HAMMER2_LOOKUP_SHARED) 2968 how_maybe |= HAMMER2_RESOLVE_SHARED; 2969 2970 parent = *parentp; 2971 *errorp = 0; 2972 2973 /* 2974 * Calculate the next index and recalculate the parent if necessary. 2975 */ 2976 if (chain) { 2977 key_beg = chain->bref.key + 2978 ((hammer2_key_t)1 << chain->bref.keybits); 2979 hammer2_chain_unlock(chain); 2980 hammer2_chain_drop(chain); 2981 2982 /* 2983 * chain invalid past this point, but we can still do a 2984 * pointer comparison w/parent. 2985 * 2986 * Any scan where the lookup returned degenerate data embedded 2987 * in the inode has an invalid index and must terminate. 2988 */ 2989 if (chain == parent) 2990 return(NULL); 2991 if (key_beg == 0 || key_beg > key_end) 2992 return(NULL); 2993 chain = NULL; 2994 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT && 2995 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) { 2996 /* 2997 * We reached the end of the iteration. 2998 */ 2999 return (NULL); 3000 } else { 3001 /* 3002 * Continue iteration with next parent unless the current 3003 * parent covers the range. 3004 * 3005 * (This also handles the case of a deleted, empty indirect 3006 * node). 3007 */ 3008 key_beg = parent->bref.key + 3009 ((hammer2_key_t)1 << parent->bref.keybits); 3010 if (key_beg == 0 || key_beg > key_end) 3011 return (NULL); 3012 parent = hammer2_chain_repparent(parentp, how_maybe); 3013 } 3014 3015 /* 3016 * And execute 3017 */ 3018 return (hammer2_chain_lookup(parentp, key_nextp, 3019 key_beg, key_end, 3020 errorp, flags)); 3021 } 3022 3023 /* 3024 * Caller wishes to iterate chains under parent, loading new chains into 3025 * chainp. Caller must initialize *chainp to NULL and *firstp to 1, and 3026 * then call hammer2_chain_scan() repeatedly until a non-zero return. 3027 * During the scan, *firstp will be set to 0 and (*chainp) will be replaced 3028 * with the returned chain for the scan. The returned *chainp will be 3029 * locked and referenced. Any prior contents will be unlocked and dropped. 3030 * 3031 * Caller should check the return value. A normal scan EOF will return 3032 * exactly HAMMER2_ERROR_EOF. Any other non-zero value indicates an 3033 * error trying to access parent data. Any error in the returned chain 3034 * must be tested separately by the caller. 3035 * 3036 * (*chainp) is dropped on each scan, but will only be set if the returned 3037 * element itself can recurse. Leaf elements are NOT resolved, loaded, or 3038 * returned via *chainp. The caller will get their bref only. 3039 * 3040 * The raw scan function is similar to lookup/next but does not seek to a key. 3041 * Blockrefs are iterated via first_bref = (parent, NULL) and 3042 * next_chain = (parent, bref). 3043 * 3044 * The passed-in parent must be locked and its data resolved. The function 3045 * nominally returns a locked and referenced *chainp != NULL for chains 3046 * the caller might need to recurse on (and will dipose of any *chainp passed 3047 * in). The caller must check the chain->bref.type either way. 3048 */ 3049 int 3050 hammer2_chain_scan(hammer2_chain_t *parent, hammer2_chain_t **chainp, 3051 hammer2_blockref_t *bref, int *firstp, 3052 int flags) 3053 { 3054 hammer2_dev_t *hmp; 3055 hammer2_blockref_t *base; 3056 hammer2_blockref_t *bref_ptr; 3057 hammer2_key_t key; 3058 hammer2_key_t next_key; 3059 hammer2_chain_t *chain = NULL; 3060 int count = 0; 3061 int how_always = HAMMER2_RESOLVE_ALWAYS; 3062 int how_maybe = HAMMER2_RESOLVE_MAYBE; 3063 int how; 3064 int generation; 3065 int maxloops = 300000; 3066 int error; 3067 3068 hmp = parent->hmp; 3069 error = 0; 3070 3071 /* 3072 * Scan flags borrowed from lookup. 3073 */ 3074 if (flags & HAMMER2_LOOKUP_ALWAYS) { 3075 how_maybe = how_always; 3076 how = HAMMER2_RESOLVE_ALWAYS; 3077 } else if (flags & HAMMER2_LOOKUP_NODATA) { 3078 how = HAMMER2_RESOLVE_NEVER; 3079 } else { 3080 how = HAMMER2_RESOLVE_MAYBE; 3081 } 3082 if (flags & HAMMER2_LOOKUP_SHARED) { 3083 how_maybe |= HAMMER2_RESOLVE_SHARED; 3084 how_always |= HAMMER2_RESOLVE_SHARED; 3085 how |= HAMMER2_RESOLVE_SHARED; 3086 } 3087 3088 /* 3089 * Calculate key to locate first/next element, unlocking the previous 3090 * element as we go. Be careful, the key calculation can overflow. 3091 * 3092 * (also reset bref to NULL) 3093 */ 3094 if (*firstp) { 3095 key = 0; 3096 *firstp = 0; 3097 } else { 3098 key = bref->key + ((hammer2_key_t)1 << bref->keybits); 3099 if ((chain = *chainp) != NULL) { 3100 *chainp = NULL; 3101 hammer2_chain_unlock(chain); 3102 hammer2_chain_drop(chain); 3103 chain = NULL; 3104 } 3105 if (key == 0) { 3106 error |= HAMMER2_ERROR_EOF; 3107 goto done; 3108 } 3109 } 3110 3111 again: 3112 if (parent->error) { 3113 error = parent->error; 3114 goto done; 3115 } 3116 if (--maxloops == 0) 3117 panic("hammer2_chain_scan: maxloops"); 3118 3119 /* 3120 * Locate the blockref array. Currently we do a fully associative 3121 * search through the array. 3122 */ 3123 switch(parent->bref.type) { 3124 case HAMMER2_BREF_TYPE_INODE: 3125 /* 3126 * An inode with embedded data has no sub-chains. 3127 * 3128 * WARNING! Bulk scan code may pass a static chain marked 3129 * as BREF_TYPE_INODE with a copy of the volume 3130 * root blockset to snapshot the volume. 3131 */ 3132 if (parent->data->ipdata.meta.op_flags & 3133 HAMMER2_OPFLAG_DIRECTDATA) { 3134 error |= HAMMER2_ERROR_EOF; 3135 goto done; 3136 } 3137 base = &parent->data->ipdata.u.blockset.blockref[0]; 3138 count = HAMMER2_SET_COUNT; 3139 break; 3140 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 3141 case HAMMER2_BREF_TYPE_INDIRECT: 3142 /* 3143 * Optimize indirect blocks in the INITIAL state to avoid 3144 * I/O. 3145 */ 3146 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 3147 base = NULL; 3148 } else { 3149 if (parent->data == NULL) 3150 panic("parent->data is NULL"); 3151 base = &parent->data->npdata[0]; 3152 } 3153 count = parent->bytes / sizeof(hammer2_blockref_t); 3154 break; 3155 case HAMMER2_BREF_TYPE_VOLUME: 3156 base = &parent->data->voldata.sroot_blockset.blockref[0]; 3157 count = HAMMER2_SET_COUNT; 3158 break; 3159 case HAMMER2_BREF_TYPE_FREEMAP: 3160 base = &parent->data->blkset.blockref[0]; 3161 count = HAMMER2_SET_COUNT; 3162 break; 3163 default: 3164 panic("hammer2_chain_scan: unrecognized blockref type: %d", 3165 parent->bref.type); 3166 base = NULL; /* safety */ 3167 count = 0; /* safety */ 3168 break; 3169 } 3170 3171 /* 3172 * Merged scan to find next candidate. 3173 * 3174 * hammer2_base_*() functions require the parent->core.live_* fields 3175 * to be synchronized. 3176 * 3177 * We need to hold the spinlock to access the block array and RB tree 3178 * and to interlock chain creation. 3179 */ 3180 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) 3181 hammer2_chain_countbrefs(parent, base, count); 3182 3183 next_key = 0; 3184 bref_ptr = NULL; 3185 hammer2_spin_ex(&parent->core.spin); 3186 chain = hammer2_combined_find(parent, base, count, 3187 &next_key, 3188 key, HAMMER2_KEY_MAX, 3189 &bref_ptr); 3190 generation = parent->core.generation; 3191 3192 /* 3193 * Exhausted parent chain, we're done. 3194 */ 3195 if (bref_ptr == NULL) { 3196 hammer2_spin_unex(&parent->core.spin); 3197 KKASSERT(chain == NULL); 3198 error |= HAMMER2_ERROR_EOF; 3199 goto done; 3200 } 3201 3202 /* 3203 * Copy into the supplied stack-based blockref. 3204 */ 3205 *bref = *bref_ptr; 3206 3207 /* 3208 * Selected from blockref or in-memory chain. 3209 */ 3210 if (chain == NULL) { 3211 switch(bref->type) { 3212 case HAMMER2_BREF_TYPE_INODE: 3213 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 3214 case HAMMER2_BREF_TYPE_INDIRECT: 3215 case HAMMER2_BREF_TYPE_VOLUME: 3216 case HAMMER2_BREF_TYPE_FREEMAP: 3217 /* 3218 * Recursion, always get the chain 3219 */ 3220 hammer2_spin_unex(&parent->core.spin); 3221 chain = hammer2_chain_get(parent, generation, 3222 bref, how); 3223 if (chain == NULL) 3224 goto again; 3225 break; 3226 default: 3227 /* 3228 * No recursion, do not waste time instantiating 3229 * a chain, just iterate using the bref. 3230 */ 3231 hammer2_spin_unex(&parent->core.spin); 3232 break; 3233 } 3234 } else { 3235 /* 3236 * Recursion or not we need the chain in order to supply 3237 * the bref. 3238 */ 3239 hammer2_chain_ref(chain); 3240 hammer2_spin_unex(&parent->core.spin); 3241 hammer2_chain_lock(chain, how); 3242 } 3243 if (chain && 3244 (bcmp(bref, &chain->bref, sizeof(*bref)) || 3245 chain->parent != parent)) { 3246 hammer2_chain_unlock(chain); 3247 hammer2_chain_drop(chain); 3248 chain = NULL; 3249 goto again; 3250 } 3251 3252 /* 3253 * Skip deleted chains (XXX cache 'i' end-of-block-array? XXX) 3254 * 3255 * NOTE: chain's key range is not relevant as there might be 3256 * one-offs within the range that are not deleted. 3257 * 3258 * NOTE: XXX this could create problems with scans used in 3259 * situations other than mount-time recovery. 3260 * 3261 * NOTE: Lookups can race delete-duplicate because 3262 * delete-duplicate does not lock the parent's core 3263 * (they just use the spinlock on the core). 3264 */ 3265 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) { 3266 hammer2_chain_unlock(chain); 3267 hammer2_chain_drop(chain); 3268 chain = NULL; 3269 3270 key = next_key; 3271 if (key == 0) { 3272 error |= HAMMER2_ERROR_EOF; 3273 goto done; 3274 } 3275 goto again; 3276 } 3277 3278 done: 3279 /* 3280 * All done, return the bref or NULL, supply chain if necessary. 3281 */ 3282 if (chain) 3283 *chainp = chain; 3284 return (error); 3285 } 3286 3287 /* 3288 * Create and return a new hammer2 system memory structure of the specified 3289 * key, type and size and insert it under (*parentp). This is a full 3290 * insertion, based on the supplied key/keybits, and may involve creating 3291 * indirect blocks and moving other chains around via delete/duplicate. 3292 * 3293 * This call can be made with parent == NULL as long as a non -1 methods 3294 * is supplied. hmp must also be supplied in this situation (otherwise 3295 * hmp is extracted from the supplied parent). The chain will be detached 3296 * from the topology. A later call with both parent and chain can be made 3297 * to attach it. 3298 * 3299 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (*parentp) TO THE INSERTION 3300 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING 3301 * FULL. This typically means that the caller is creating the chain after 3302 * doing a hammer2_chain_lookup(). 3303 * 3304 * (*parentp) must be exclusive locked and may be replaced on return 3305 * depending on how much work the function had to do. 3306 * 3307 * (*parentp) must not be errored or this function will assert. 3308 * 3309 * (*chainp) usually starts out NULL and returns the newly created chain, 3310 * but if the caller desires the caller may allocate a disconnected chain 3311 * and pass it in instead. 3312 * 3313 * This function should NOT be used to insert INDIRECT blocks. It is 3314 * typically used to create/insert inodes and data blocks. 3315 * 3316 * Caller must pass-in an exclusively locked parent the new chain is to 3317 * be inserted under, and optionally pass-in a disconnected, exclusively 3318 * locked chain to insert (else we create a new chain). The function will 3319 * adjust (*parentp) as necessary, create or connect the chain, and 3320 * return an exclusively locked chain in *chainp. 3321 * 3322 * When creating a PFSROOT inode under the super-root, pmp is typically NULL 3323 * and will be reassigned. 3324 * 3325 * NOTE: returns HAMMER_ERROR_* flags 3326 */ 3327 int 3328 hammer2_chain_create(hammer2_chain_t **parentp, hammer2_chain_t **chainp, 3329 hammer2_dev_t *hmp, hammer2_pfs_t *pmp, int methods, 3330 hammer2_key_t key, int keybits, int type, size_t bytes, 3331 hammer2_tid_t mtid, hammer2_off_t dedup_off, int flags) 3332 { 3333 hammer2_chain_t *chain; 3334 hammer2_chain_t *parent; 3335 hammer2_blockref_t *base; 3336 hammer2_blockref_t dummy; 3337 int allocated = 0; 3338 int error = 0; 3339 int count; 3340 int maxloops = 300000; 3341 3342 /* 3343 * Topology may be crossing a PFS boundary. 3344 */ 3345 parent = *parentp; 3346 if (parent) { 3347 KKASSERT(hammer2_mtx_owned(&parent->lock)); 3348 KKASSERT(parent->error == 0); 3349 hmp = parent->hmp; 3350 } 3351 chain = *chainp; 3352 3353 if (chain == NULL) { 3354 /* 3355 * First allocate media space and construct the dummy bref, 3356 * then allocate the in-memory chain structure. Set the 3357 * INITIAL flag for fresh chains which do not have embedded 3358 * data. 3359 */ 3360 bzero(&dummy, sizeof(dummy)); 3361 dummy.type = type; 3362 dummy.key = key; 3363 dummy.keybits = keybits; 3364 dummy.data_off = hammer2_getradix(bytes); 3365 3366 /* 3367 * Inherit methods from parent by default. Primarily used 3368 * for BREF_TYPE_DATA. Non-data types *must* be set to 3369 * a non-NONE check algorithm. 3370 */ 3371 if (methods == HAMMER2_METH_DEFAULT) 3372 dummy.methods = parent->bref.methods; 3373 else 3374 dummy.methods = (uint8_t)methods; 3375 3376 if (type != HAMMER2_BREF_TYPE_DATA && 3377 HAMMER2_DEC_CHECK(dummy.methods) == HAMMER2_CHECK_NONE) { 3378 dummy.methods |= 3379 HAMMER2_ENC_CHECK(HAMMER2_CHECK_DEFAULT); 3380 } 3381 3382 chain = hammer2_chain_alloc(hmp, pmp, &dummy); 3383 3384 /* 3385 * Lock the chain manually, chain_lock will load the chain 3386 * which we do NOT want to do. (note: chain->refs is set 3387 * to 1 by chain_alloc() for us, but lockcnt is not). 3388 */ 3389 chain->lockcnt = 1; 3390 hammer2_mtx_ex(&chain->lock); 3391 allocated = 1; 3392 3393 /* 3394 * Set INITIAL to optimize I/O. The flag will generally be 3395 * processed when we call hammer2_chain_modify(). 3396 */ 3397 switch(type) { 3398 case HAMMER2_BREF_TYPE_VOLUME: 3399 case HAMMER2_BREF_TYPE_FREEMAP: 3400 panic("hammer2_chain_create: called with volume type"); 3401 break; 3402 case HAMMER2_BREF_TYPE_INDIRECT: 3403 panic("hammer2_chain_create: cannot be used to" 3404 "create indirect block"); 3405 break; 3406 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 3407 panic("hammer2_chain_create: cannot be used to" 3408 "create freemap root or node"); 3409 break; 3410 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 3411 KKASSERT(bytes == sizeof(chain->data->bmdata)); 3412 /* fall through */ 3413 case HAMMER2_BREF_TYPE_DIRENT: 3414 case HAMMER2_BREF_TYPE_INODE: 3415 case HAMMER2_BREF_TYPE_DATA: 3416 default: 3417 /* 3418 * leave chain->data NULL, set INITIAL 3419 */ 3420 KKASSERT(chain->data == NULL); 3421 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 3422 break; 3423 } 3424 } else { 3425 /* 3426 * We are reattaching a previously deleted chain, possibly 3427 * under a new parent and possibly with a new key/keybits. 3428 * The chain does not have to be in a modified state. The 3429 * UPDATE flag will be set later on in this routine. 3430 * 3431 * Do NOT mess with the current state of the INITIAL flag. 3432 */ 3433 chain->bref.key = key; 3434 chain->bref.keybits = keybits; 3435 if (chain->flags & HAMMER2_CHAIN_DELETED) 3436 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DELETED); 3437 KKASSERT(chain->parent == NULL); 3438 } 3439 3440 /* 3441 * Set the appropriate bref flag if requested. 3442 * 3443 * NOTE! Callers can call this function to move chains without 3444 * knowing about special flags, so don't clear bref flags 3445 * here! 3446 */ 3447 if (flags & HAMMER2_INSERT_PFSROOT) 3448 chain->bref.flags |= HAMMER2_BREF_FLAG_PFSROOT; 3449 3450 if (parent == NULL) 3451 goto skip; 3452 3453 /* 3454 * Calculate how many entries we have in the blockref array and 3455 * determine if an indirect block is required when inserting into 3456 * the parent. 3457 */ 3458 again: 3459 if (--maxloops == 0) 3460 panic("hammer2_chain_create: maxloops"); 3461 3462 switch(parent->bref.type) { 3463 case HAMMER2_BREF_TYPE_INODE: 3464 if ((parent->data->ipdata.meta.op_flags & 3465 HAMMER2_OPFLAG_DIRECTDATA) != 0) { 3466 kprintf("hammer2: parent set for direct-data! " 3467 "pkey=%016jx ckey=%016jx\n", 3468 parent->bref.key, 3469 chain->bref.key); 3470 } 3471 KKASSERT((parent->data->ipdata.meta.op_flags & 3472 HAMMER2_OPFLAG_DIRECTDATA) == 0); 3473 KKASSERT(parent->data != NULL); 3474 base = &parent->data->ipdata.u.blockset.blockref[0]; 3475 count = HAMMER2_SET_COUNT; 3476 break; 3477 case HAMMER2_BREF_TYPE_INDIRECT: 3478 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 3479 if (parent->flags & HAMMER2_CHAIN_INITIAL) 3480 base = NULL; 3481 else 3482 base = &parent->data->npdata[0]; 3483 count = parent->bytes / sizeof(hammer2_blockref_t); 3484 break; 3485 case HAMMER2_BREF_TYPE_VOLUME: 3486 KKASSERT(parent->data != NULL); 3487 base = &parent->data->voldata.sroot_blockset.blockref[0]; 3488 count = HAMMER2_SET_COUNT; 3489 break; 3490 case HAMMER2_BREF_TYPE_FREEMAP: 3491 KKASSERT(parent->data != NULL); 3492 base = &parent->data->blkset.blockref[0]; 3493 count = HAMMER2_SET_COUNT; 3494 break; 3495 default: 3496 panic("hammer2_chain_create: unrecognized blockref type: %d", 3497 parent->bref.type); 3498 base = NULL; 3499 count = 0; 3500 break; 3501 } 3502 3503 /* 3504 * Make sure we've counted the brefs 3505 */ 3506 if ((parent->flags & HAMMER2_CHAIN_COUNTEDBREFS) == 0) 3507 hammer2_chain_countbrefs(parent, base, count); 3508 3509 KASSERT(parent->core.live_count >= 0 && 3510 parent->core.live_count <= count, 3511 ("bad live_count %d/%d (%02x, %d)", 3512 parent->core.live_count, count, 3513 parent->bref.type, parent->bytes)); 3514 3515 /* 3516 * If no free blockref could be found we must create an indirect 3517 * block and move a number of blockrefs into it. With the parent 3518 * locked we can safely lock each child in order to delete+duplicate 3519 * it without causing a deadlock. 3520 * 3521 * This may return the new indirect block or the old parent depending 3522 * on where the key falls. NULL is returned on error. 3523 */ 3524 if (parent->core.live_count == count) { 3525 hammer2_chain_t *nparent; 3526 3527 KKASSERT((flags & HAMMER2_INSERT_SAMEPARENT) == 0); 3528 3529 nparent = hammer2_chain_create_indirect(parent, key, keybits, 3530 mtid, type, &error); 3531 if (nparent == NULL) { 3532 if (allocated) 3533 hammer2_chain_drop(chain); 3534 chain = NULL; 3535 goto done; 3536 } 3537 if (parent != nparent) { 3538 hammer2_chain_unlock(parent); 3539 hammer2_chain_drop(parent); 3540 parent = *parentp = nparent; 3541 } 3542 goto again; 3543 } 3544 3545 /* 3546 * fall through if parent, or skip to here if no parent. 3547 */ 3548 skip: 3549 if (chain->flags & HAMMER2_CHAIN_DELETED) 3550 kprintf("Inserting deleted chain @%016jx\n", 3551 chain->bref.key); 3552 3553 /* 3554 * Link the chain into its parent. 3555 */ 3556 if (chain->parent != NULL) 3557 panic("hammer2: hammer2_chain_create: chain already connected"); 3558 KKASSERT(chain->parent == NULL); 3559 if (parent) { 3560 KKASSERT(parent->core.live_count < count); 3561 hammer2_chain_insert(parent, chain, 3562 HAMMER2_CHAIN_INSERT_SPIN | 3563 HAMMER2_CHAIN_INSERT_LIVE, 3564 0); 3565 } 3566 3567 if (allocated) { 3568 /* 3569 * Mark the newly created chain modified. This will cause 3570 * UPDATE to be set and process the INITIAL flag. 3571 * 3572 * Device buffers are not instantiated for DATA elements 3573 * as these are handled by logical buffers. 3574 * 3575 * Indirect and freemap node indirect blocks are handled 3576 * by hammer2_chain_create_indirect() and not by this 3577 * function. 3578 * 3579 * Data for all other bref types is expected to be 3580 * instantiated (INODE, LEAF). 3581 */ 3582 switch(chain->bref.type) { 3583 case HAMMER2_BREF_TYPE_DATA: 3584 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 3585 case HAMMER2_BREF_TYPE_DIRENT: 3586 case HAMMER2_BREF_TYPE_INODE: 3587 error = hammer2_chain_modify(chain, mtid, dedup_off, 3588 HAMMER2_MODIFY_OPTDATA); 3589 break; 3590 default: 3591 /* 3592 * Remaining types are not supported by this function. 3593 * In particular, INDIRECT and LEAF_NODE types are 3594 * handled by create_indirect(). 3595 */ 3596 panic("hammer2_chain_create: bad type: %d", 3597 chain->bref.type); 3598 /* NOT REACHED */ 3599 break; 3600 } 3601 } else { 3602 /* 3603 * When reconnecting a chain we must set UPDATE and 3604 * setflush so the flush recognizes that it must update 3605 * the bref in the parent. 3606 */ 3607 if ((chain->flags & HAMMER2_CHAIN_UPDATE) == 0) 3608 atomic_set_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 3609 } 3610 3611 /* 3612 * We must setflush(parent) to ensure that it recurses through to 3613 * chain. setflush(chain) might not work because ONFLUSH is possibly 3614 * already set in the chain (so it won't recurse up to set it in the 3615 * parent). 3616 */ 3617 if (parent) 3618 hammer2_chain_setflush(parent); 3619 3620 done: 3621 *chainp = chain; 3622 3623 return (error); 3624 } 3625 3626 /* 3627 * Move the chain from its old parent to a new parent. The chain must have 3628 * already been deleted or already disconnected (or never associated) with 3629 * a parent. The chain is reassociated with the new parent and the deleted 3630 * flag will be cleared (no longer deleted). The chain's modification state 3631 * is not altered. 3632 * 3633 * THE CALLER MUST HAVE ALREADY PROPERLY SEEKED (parent) TO THE INSERTION 3634 * POINT SANS ANY REQUIRED INDIRECT BLOCK CREATIONS DUE TO THE ARRAY BEING 3635 * FULL. This typically means that the caller is creating the chain after 3636 * doing a hammer2_chain_lookup(). 3637 * 3638 * Neither (parent) or (chain) can be errored. 3639 * 3640 * If (parent) is non-NULL then the chain is inserted under the parent. 3641 * 3642 * If (parent) is NULL then the newly duplicated chain is not inserted 3643 * anywhere, similar to if it had just been chain_alloc()'d (suitable for 3644 * passing into hammer2_chain_create() after this function returns). 3645 * 3646 * WARNING! This function calls create which means it can insert indirect 3647 * blocks. This can cause other unrelated chains in the parent to 3648 * be moved to a newly inserted indirect block in addition to the 3649 * specific chain. 3650 */ 3651 void 3652 hammer2_chain_rename(hammer2_chain_t **parentp, hammer2_chain_t *chain, 3653 hammer2_tid_t mtid, int flags) 3654 { 3655 hammer2_blockref_t *bref; 3656 hammer2_dev_t *hmp; 3657 hammer2_chain_t *parent; 3658 3659 /* 3660 * WARNING! We should never resolve DATA to device buffers 3661 * (XXX allow it if the caller did?), and since 3662 * we currently do not have the logical buffer cache 3663 * buffer in-hand to fix its cached physical offset 3664 * we also force the modify code to not COW it. XXX 3665 * 3666 * NOTE! We allow error'd chains to be renamed. The bref itself 3667 * is good and can be renamed. The content, however, may 3668 * be inaccessible. 3669 */ 3670 hmp = chain->hmp; 3671 KKASSERT(chain->parent == NULL); 3672 /*KKASSERT(chain->error == 0); allow */ 3673 bref = &chain->bref; 3674 3675 /* 3676 * If parent is not NULL the duplicated chain will be entered under 3677 * the parent and the UPDATE bit set to tell flush to update 3678 * the blockref. 3679 * 3680 * We must setflush(parent) to ensure that it recurses through to 3681 * chain. setflush(chain) might not work because ONFLUSH is possibly 3682 * already set in the chain (so it won't recurse up to set it in the 3683 * parent). 3684 * 3685 * Having both chains locked is extremely important for atomicy. 3686 */ 3687 if (parentp && (parent = *parentp) != NULL) { 3688 KKASSERT(hammer2_mtx_owned(&parent->lock)); 3689 KKASSERT(parent->refs > 0); 3690 KKASSERT(parent->error == 0); 3691 3692 hammer2_chain_create(parentp, &chain, NULL, chain->pmp, 3693 HAMMER2_METH_DEFAULT, 3694 bref->key, bref->keybits, bref->type, 3695 chain->bytes, mtid, 0, flags); 3696 KKASSERT(chain->flags & HAMMER2_CHAIN_UPDATE); 3697 hammer2_chain_setflush(*parentp); 3698 } 3699 } 3700 3701 /* 3702 * This works in tandem with delete_obref() to install a blockref in 3703 * (typically) an indirect block that is associated with the chain being 3704 * moved to *parentp. 3705 * 3706 * The reason we need this function is that the caller needs to maintain 3707 * the blockref as it was, and not generate a new blockref for what might 3708 * be a modified chain. Otherwise stuff will leak into the flush that 3709 * the flush code's FLUSH_INODE_STOP flag is unable to catch. 3710 * 3711 * It is EXTREMELY important that we properly set CHAIN_BLKMAPUPD and 3712 * CHAIN_UPDATE. We must set BLKMAPUPD if the bref does not match, and 3713 * we must clear CHAIN_UPDATE (that was likely set by the chain_rename) if 3714 * it does. Otherwise we can end up in a situation where H2 is unable to 3715 * clean up the in-memory chain topology. 3716 * 3717 * The reason for this is that flushes do not generally flush through 3718 * BREF_TYPE_INODE chains and depend on a hammer2_inode_t queued to syncq 3719 * or sideq to properly flush and dispose of the related inode chain's flags. 3720 * Situations where the inode is not actually modified by the frontend, 3721 * but where we have to move the related chains around as we insert or cleanup 3722 * indirect blocks, can leave us with a 'dirty' (non-disposable) in-memory 3723 * inode chain that does not have a hammer2_inode_t associated with it. 3724 */ 3725 static void 3726 hammer2_chain_rename_obref(hammer2_chain_t **parentp, hammer2_chain_t *chain, 3727 hammer2_tid_t mtid, int flags, 3728 hammer2_blockref_t *obref) 3729 { 3730 hammer2_chain_rename(parentp, chain, mtid, flags); 3731 3732 if (obref->type != HAMMER2_BREF_TYPE_EMPTY) { 3733 hammer2_blockref_t *tbase; 3734 int tcount; 3735 3736 KKASSERT((chain->flags & HAMMER2_CHAIN_BLKMAPPED) == 0); 3737 hammer2_chain_modify(*parentp, mtid, 0, 0); 3738 tbase = hammer2_chain_base_and_count(*parentp, &tcount); 3739 hammer2_base_insert(*parentp, tbase, tcount, chain, obref); 3740 if (bcmp(obref, &chain->bref, sizeof(chain->bref))) { 3741 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPUPD | 3742 HAMMER2_CHAIN_UPDATE); 3743 } else { 3744 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_UPDATE); 3745 } 3746 } 3747 } 3748 3749 /* 3750 * Helper function for deleting chains. 3751 * 3752 * The chain is removed from the live view (the RBTREE) as well as the parent's 3753 * blockmap. Both chain and its parent must be locked. 3754 * 3755 * parent may not be errored. chain can be errored. 3756 */ 3757 static int 3758 _hammer2_chain_delete_helper(hammer2_chain_t *parent, hammer2_chain_t *chain, 3759 hammer2_tid_t mtid, int flags, 3760 hammer2_blockref_t *obref) 3761 { 3762 hammer2_dev_t *hmp; 3763 int error = 0; 3764 3765 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0); 3766 KKASSERT(chain->parent == parent); 3767 hmp = chain->hmp; 3768 3769 if (chain->flags & HAMMER2_CHAIN_BLKMAPPED) { 3770 /* 3771 * Chain is blockmapped, so there must be a parent. 3772 * Atomically remove the chain from the parent and remove 3773 * the blockmap entry. The parent must be set modified 3774 * to remove the blockmap entry. 3775 */ 3776 hammer2_blockref_t *base; 3777 int count; 3778 3779 KKASSERT(parent != NULL); 3780 KKASSERT(parent->error == 0); 3781 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0); 3782 error = hammer2_chain_modify(parent, mtid, 0, 0); 3783 if (error) 3784 goto done; 3785 3786 /* 3787 * Calculate blockmap pointer 3788 */ 3789 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE); 3790 hammer2_spin_ex(&chain->core.spin); 3791 hammer2_spin_ex(&parent->core.spin); 3792 3793 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED); 3794 atomic_add_int(&parent->core.live_count, -1); 3795 ++parent->core.generation; 3796 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain); 3797 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 3798 --parent->core.chain_count; 3799 chain->parent = NULL; 3800 3801 switch(parent->bref.type) { 3802 case HAMMER2_BREF_TYPE_INODE: 3803 /* 3804 * Access the inode's block array. However, there 3805 * is no block array if the inode is flagged 3806 * DIRECTDATA. 3807 */ 3808 if (parent->data && 3809 (parent->data->ipdata.meta.op_flags & 3810 HAMMER2_OPFLAG_DIRECTDATA) == 0) { 3811 base = 3812 &parent->data->ipdata.u.blockset.blockref[0]; 3813 } else { 3814 base = NULL; 3815 } 3816 count = HAMMER2_SET_COUNT; 3817 break; 3818 case HAMMER2_BREF_TYPE_INDIRECT: 3819 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 3820 if (parent->data) 3821 base = &parent->data->npdata[0]; 3822 else 3823 base = NULL; 3824 count = parent->bytes / sizeof(hammer2_blockref_t); 3825 break; 3826 case HAMMER2_BREF_TYPE_VOLUME: 3827 base = &parent->data->voldata. 3828 sroot_blockset.blockref[0]; 3829 count = HAMMER2_SET_COUNT; 3830 break; 3831 case HAMMER2_BREF_TYPE_FREEMAP: 3832 base = &parent->data->blkset.blockref[0]; 3833 count = HAMMER2_SET_COUNT; 3834 break; 3835 default: 3836 base = NULL; 3837 count = 0; 3838 panic("_hammer2_chain_delete_helper: " 3839 "unrecognized blockref type: %d", 3840 parent->bref.type); 3841 break; 3842 } 3843 3844 /* 3845 * delete blockmapped chain from its parent. 3846 * 3847 * The parent is not affected by any statistics in chain 3848 * which are pending synchronization. That is, there is 3849 * nothing to undo in the parent since they have not yet 3850 * been incorporated into the parent. 3851 * 3852 * The parent is affected by statistics stored in inodes. 3853 * Those have already been synchronized, so they must be 3854 * undone. XXX split update possible w/delete in middle? 3855 */ 3856 if (base) { 3857 hammer2_base_delete(parent, base, count, chain, obref); 3858 } 3859 hammer2_spin_unex(&parent->core.spin); 3860 hammer2_spin_unex(&chain->core.spin); 3861 } else if (chain->flags & HAMMER2_CHAIN_ONRBTREE) { 3862 /* 3863 * Chain is not blockmapped but a parent is present. 3864 * Atomically remove the chain from the parent. There is 3865 * no blockmap entry to remove. 3866 * 3867 * Because chain was associated with a parent but not 3868 * synchronized, the chain's *_count_up fields contain 3869 * inode adjustment statistics which must be undone. 3870 */ 3871 hammer2_spin_ex(&chain->core.spin); 3872 hammer2_spin_ex(&parent->core.spin); 3873 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED); 3874 atomic_add_int(&parent->core.live_count, -1); 3875 ++parent->core.generation; 3876 RB_REMOVE(hammer2_chain_tree, &parent->core.rbtree, chain); 3877 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 3878 --parent->core.chain_count; 3879 chain->parent = NULL; 3880 hammer2_spin_unex(&parent->core.spin); 3881 hammer2_spin_unex(&chain->core.spin); 3882 } else { 3883 /* 3884 * Chain is not blockmapped and has no parent. This 3885 * is a degenerate case. 3886 */ 3887 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED); 3888 } 3889 done: 3890 return error; 3891 } 3892 3893 /* 3894 * Create an indirect block that covers one or more of the elements in the 3895 * current parent. Either returns the existing parent with no locking or 3896 * ref changes or returns the new indirect block locked and referenced 3897 * and leaving the original parent lock/ref intact as well. 3898 * 3899 * If an error occurs, NULL is returned and *errorp is set to the H2 error. 3900 * 3901 * The returned chain depends on where the specified key falls. 3902 * 3903 * The key/keybits for the indirect mode only needs to follow three rules: 3904 * 3905 * (1) That all elements underneath it fit within its key space and 3906 * 3907 * (2) That all elements outside it are outside its key space. 3908 * 3909 * (3) When creating the new indirect block any elements in the current 3910 * parent that fit within the new indirect block's keyspace must be 3911 * moved into the new indirect block. 3912 * 3913 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider 3914 * keyspace the the current parent, but lookup/iteration rules will 3915 * ensure (and must ensure) that rule (2) for all parents leading up 3916 * to the nearest inode or the root volume header is adhered to. This 3917 * is accomplished by always recursing through matching keyspaces in 3918 * the hammer2_chain_lookup() and hammer2_chain_next() API. 3919 * 3920 * The current implementation calculates the current worst-case keyspace by 3921 * iterating the current parent and then divides it into two halves, choosing 3922 * whichever half has the most elements (not necessarily the half containing 3923 * the requested key). 3924 * 3925 * We can also opt to use the half with the least number of elements. This 3926 * causes lower-numbered keys (aka logical file offsets) to recurse through 3927 * fewer indirect blocks and higher-numbered keys to recurse through more. 3928 * This also has the risk of not moving enough elements to the new indirect 3929 * block and being forced to create several indirect blocks before the element 3930 * can be inserted. 3931 * 3932 * Must be called with an exclusively locked parent. 3933 * 3934 * NOTE: *errorp set to HAMMER_ERROR_* flags 3935 */ 3936 static int hammer2_chain_indkey_freemap(hammer2_chain_t *parent, 3937 hammer2_key_t *keyp, int keybits, 3938 hammer2_blockref_t *base, int count); 3939 static int hammer2_chain_indkey_file(hammer2_chain_t *parent, 3940 hammer2_key_t *keyp, int keybits, 3941 hammer2_blockref_t *base, int count, 3942 int ncount); 3943 static int hammer2_chain_indkey_dir(hammer2_chain_t *parent, 3944 hammer2_key_t *keyp, int keybits, 3945 hammer2_blockref_t *base, int count, 3946 int ncount); 3947 static 3948 hammer2_chain_t * 3949 hammer2_chain_create_indirect(hammer2_chain_t *parent, 3950 hammer2_key_t create_key, int create_bits, 3951 hammer2_tid_t mtid, int for_type, int *errorp) 3952 { 3953 hammer2_dev_t *hmp; 3954 hammer2_blockref_t *base; 3955 hammer2_blockref_t *bref; 3956 hammer2_blockref_t bsave; 3957 hammer2_blockref_t dummy; 3958 hammer2_chain_t *chain; 3959 hammer2_chain_t *ichain; 3960 hammer2_key_t key = create_key; 3961 hammer2_key_t key_beg; 3962 hammer2_key_t key_end; 3963 hammer2_key_t key_next; 3964 int keybits = create_bits; 3965 int count; 3966 int ncount; 3967 int nbytes; 3968 int loops; 3969 int error; 3970 int reason; 3971 int generation; 3972 int maxloops = 300000; 3973 3974 /* 3975 * Calculate the base blockref pointer or NULL if the chain 3976 * is known to be empty. We need to calculate the array count 3977 * for RB lookups either way. 3978 */ 3979 hmp = parent->hmp; 3980 KKASSERT(hammer2_mtx_owned(&parent->lock)); 3981 3982 /* 3983 * Pre-modify the parent now to avoid having to deal with error 3984 * processing if we tried to later (in the middle of our loop). 3985 * 3986 * We are going to be moving bref's around, the indirect blocks 3987 * cannot be in an initial state. Do not pass MODIFY_OPTDATA. 3988 */ 3989 *errorp = hammer2_chain_modify(parent, mtid, 0, 0); 3990 if (*errorp) { 3991 kprintf("hammer2_chain_create_indirect: error %08x %s\n", 3992 *errorp, hammer2_error_str(*errorp)); 3993 return NULL; 3994 } 3995 KKASSERT((parent->flags & HAMMER2_CHAIN_INITIAL) == 0); 3996 3997 /*hammer2_chain_modify(&parent, HAMMER2_MODIFY_OPTDATA);*/ 3998 base = hammer2_chain_base_and_count(parent, &count); 3999 4000 /* 4001 * How big should our new indirect block be? It has to be at least 4002 * as large as its parent for splits to work properly. 4003 * 4004 * The freemap uses a specific indirect block size. The number of 4005 * levels are built dynamically and ultimately depend on the size 4006 * volume. Because freemap blocks are taken from the reserved areas 4007 * of the volume our goal is efficiency (fewer levels) and not so 4008 * much to save disk space. 4009 * 4010 * The first indirect block level for a directory usually uses 4011 * HAMMER2_IND_BYTES_MIN (4KB = 32 directory entries). Due to 4012 * the hash mechanism, this typically gives us a nominal 4013 * 32 * 4 entries with one level of indirection. 4014 * 4015 * We use HAMMER2_IND_BYTES_NOM (16KB = 128 blockrefs) for FILE 4016 * indirect blocks. The initial 4 entries in the inode gives us 4017 * 256KB. Up to 4 indirect blocks gives us 32MB. Three levels 4018 * of indirection gives us 137GB, and so forth. H2 can support 4019 * huge file sizes but they are not typical, so we try to stick 4020 * with compactness and do not use a larger indirect block size. 4021 * 4022 * We could use 64KB (PBUFSIZE), giving us 512 blockrefs, but 4023 * due to the way indirect blocks are created this usually winds 4024 * up being extremely inefficient for small files. Even though 4025 * 16KB requires more levels of indirection for very large files, 4026 * the 16KB records can be ganged together into 64KB DIOs. 4027 */ 4028 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE || 4029 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) { 4030 nbytes = HAMMER2_FREEMAP_LEVELN_PSIZE; 4031 } else if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) { 4032 if (parent->data->ipdata.meta.type == 4033 HAMMER2_OBJTYPE_DIRECTORY) 4034 nbytes = HAMMER2_IND_BYTES_MIN; /* 4KB = 32 entries */ 4035 else 4036 nbytes = HAMMER2_IND_BYTES_NOM; /* 16KB = ~8MB file */ 4037 4038 } else { 4039 nbytes = HAMMER2_IND_BYTES_NOM; 4040 } 4041 if (nbytes < count * sizeof(hammer2_blockref_t)) { 4042 KKASSERT(for_type != HAMMER2_BREF_TYPE_FREEMAP_NODE && 4043 for_type != HAMMER2_BREF_TYPE_FREEMAP_LEAF); 4044 nbytes = count * sizeof(hammer2_blockref_t); 4045 } 4046 ncount = nbytes / sizeof(hammer2_blockref_t); 4047 4048 /* 4049 * When creating an indirect block for a freemap node or leaf 4050 * the key/keybits must be fitted to static radix levels because 4051 * particular radix levels use particular reserved blocks in the 4052 * related zone. 4053 * 4054 * This routine calculates the key/radix of the indirect block 4055 * we need to create, and whether it is on the high-side or the 4056 * low-side. 4057 */ 4058 switch(for_type) { 4059 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 4060 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 4061 keybits = hammer2_chain_indkey_freemap(parent, &key, keybits, 4062 base, count); 4063 break; 4064 case HAMMER2_BREF_TYPE_DATA: 4065 keybits = hammer2_chain_indkey_file(parent, &key, keybits, 4066 base, count, ncount); 4067 break; 4068 case HAMMER2_BREF_TYPE_DIRENT: 4069 case HAMMER2_BREF_TYPE_INODE: 4070 keybits = hammer2_chain_indkey_dir(parent, &key, keybits, 4071 base, count, ncount); 4072 break; 4073 default: 4074 panic("illegal indirect block for bref type %d", for_type); 4075 break; 4076 } 4077 4078 /* 4079 * Normalize the key for the radix being represented, keeping the 4080 * high bits and throwing away the low bits. 4081 */ 4082 key &= ~(((hammer2_key_t)1 << keybits) - 1); 4083 4084 /* 4085 * Ok, create our new indirect block 4086 */ 4087 bzero(&dummy, sizeof(dummy)); 4088 if (for_type == HAMMER2_BREF_TYPE_FREEMAP_NODE || 4089 for_type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) { 4090 dummy.type = HAMMER2_BREF_TYPE_FREEMAP_NODE; 4091 } else { 4092 dummy.type = HAMMER2_BREF_TYPE_INDIRECT; 4093 } 4094 dummy.key = key; 4095 dummy.keybits = keybits; 4096 dummy.data_off = hammer2_getradix(nbytes); 4097 dummy.methods = 4098 HAMMER2_ENC_CHECK(HAMMER2_DEC_CHECK(parent->bref.methods)) | 4099 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE); 4100 4101 ichain = hammer2_chain_alloc(hmp, parent->pmp, &dummy); 4102 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL); 4103 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE); 4104 /* ichain has one ref at this point */ 4105 4106 /* 4107 * We have to mark it modified to allocate its block, but use 4108 * OPTDATA to allow it to remain in the INITIAL state. Otherwise 4109 * it won't be acted upon by the flush code. 4110 * 4111 * XXX remove OPTDATA, we need a fully initialized indirect block to 4112 * be able to move the original blockref. 4113 */ 4114 *errorp = hammer2_chain_modify(ichain, mtid, 0, 0); 4115 if (*errorp) { 4116 kprintf("hammer2_chain_create_indirect: error %08x %s\n", 4117 *errorp, hammer2_error_str(*errorp)); 4118 hammer2_chain_unlock(ichain); 4119 hammer2_chain_drop(ichain); 4120 return NULL; 4121 } 4122 KKASSERT((ichain->flags & HAMMER2_CHAIN_INITIAL) == 0); 4123 4124 /* 4125 * Iterate the original parent and move the matching brefs into 4126 * the new indirect block. 4127 * 4128 * XXX handle flushes. 4129 */ 4130 key_beg = 0; 4131 key_end = HAMMER2_KEY_MAX; 4132 key_next = 0; /* avoid gcc warnings */ 4133 hammer2_spin_ex(&parent->core.spin); 4134 loops = 0; 4135 reason = 0; 4136 4137 for (;;) { 4138 /* 4139 * Parent may have been modified, relocating its block array. 4140 * Reload the base pointer. 4141 */ 4142 base = hammer2_chain_base_and_count(parent, &count); 4143 4144 if (++loops > 100000) { 4145 hammer2_spin_unex(&parent->core.spin); 4146 panic("excessive loops r=%d p=%p base/count %p:%d %016jx\n", 4147 reason, parent, base, count, key_next); 4148 } 4149 4150 /* 4151 * NOTE: spinlock stays intact, returned chain (if not NULL) 4152 * is not referenced or locked which means that we 4153 * cannot safely check its flagged / deletion status 4154 * until we lock it. 4155 */ 4156 chain = hammer2_combined_find(parent, base, count, 4157 &key_next, 4158 key_beg, key_end, 4159 &bref); 4160 generation = parent->core.generation; 4161 if (bref == NULL) 4162 break; 4163 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4164 4165 /* 4166 * Skip keys that are not within the key/radix of the new 4167 * indirect block. They stay in the parent. 4168 */ 4169 if (rounddown2(key ^ bref->key, (hammer2_key_t)1 << keybits) != 0) { 4170 goto next_key_spinlocked; 4171 } 4172 4173 /* 4174 * Load the new indirect block by acquiring the related 4175 * chains (potentially from media as it might not be 4176 * in-memory). Then move it to the new parent (ichain). 4177 * 4178 * chain is referenced but not locked. We must lock the 4179 * chain to obtain definitive state. 4180 */ 4181 bsave = *bref; 4182 if (chain) { 4183 /* 4184 * Use chain already present in the RBTREE 4185 */ 4186 hammer2_chain_ref(chain); 4187 hammer2_spin_unex(&parent->core.spin); 4188 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER); 4189 } else { 4190 /* 4191 * Get chain for blockref element. _get returns NULL 4192 * on insertion race. 4193 */ 4194 hammer2_spin_unex(&parent->core.spin); 4195 chain = hammer2_chain_get(parent, generation, &bsave, 4196 HAMMER2_RESOLVE_NEVER); 4197 if (chain == NULL) { 4198 reason = 1; 4199 hammer2_spin_ex(&parent->core.spin); 4200 continue; 4201 } 4202 } 4203 4204 /* 4205 * This is always live so if the chain has been deleted 4206 * we raced someone and we have to retry. 4207 * 4208 * NOTE: Lookups can race delete-duplicate because 4209 * delete-duplicate does not lock the parent's core 4210 * (they just use the spinlock on the core). 4211 * 4212 * (note reversed logic for this one) 4213 */ 4214 if (bcmp(&bsave, &chain->bref, sizeof(bsave)) || 4215 chain->parent != parent || 4216 (chain->flags & HAMMER2_CHAIN_DELETED)) { 4217 hammer2_chain_unlock(chain); 4218 hammer2_chain_drop(chain); 4219 if (hammer2_debug & 0x0040) { 4220 kprintf("LOST PARENT RETRY " 4221 "RETRY (%p,%p)->%p %08x\n", 4222 parent, chain->parent, chain, chain->flags); 4223 } 4224 hammer2_spin_ex(&parent->core.spin); 4225 continue; 4226 } 4227 4228 /* 4229 * Shift the chain to the indirect block. 4230 * 4231 * WARNING! No reason for us to load chain data, pass NOSTATS 4232 * to prevent delete/insert from trying to access 4233 * inode stats (and thus asserting if there is no 4234 * chain->data loaded). 4235 * 4236 * WARNING! The (parent, chain) deletion may modify the parent 4237 * and invalidate the base pointer. 4238 * 4239 * WARNING! Parent must already be marked modified, so we 4240 * can assume that chain_delete always suceeds. 4241 * 4242 * WARNING! hammer2_chain_repchange() does not have to be 4243 * called (and doesn't work anyway because we are 4244 * only doing a partial shift). A recursion that is 4245 * in-progress can continue at the current parent 4246 * and will be able to properly find its next key. 4247 */ 4248 error = hammer2_chain_delete_obref(parent, chain, mtid, 0, 4249 &bsave); 4250 KKASSERT(error == 0); 4251 hammer2_chain_rename_obref(&ichain, chain, mtid, 0, &bsave); 4252 hammer2_chain_unlock(chain); 4253 hammer2_chain_drop(chain); 4254 KKASSERT(parent->refs > 0); 4255 chain = NULL; 4256 base = NULL; /* safety */ 4257 hammer2_spin_ex(&parent->core.spin); 4258 next_key_spinlocked: 4259 if (--maxloops == 0) 4260 panic("hammer2_chain_create_indirect: maxloops"); 4261 reason = 4; 4262 if (key_next == 0 || key_next > key_end) 4263 break; 4264 key_beg = key_next; 4265 /* loop */ 4266 } 4267 hammer2_spin_unex(&parent->core.spin); 4268 4269 /* 4270 * Insert the new indirect block into the parent now that we've 4271 * cleared out some entries in the parent. We calculated a good 4272 * insertion index in the loop above (ichain->index). 4273 * 4274 * We don't have to set UPDATE here because we mark ichain 4275 * modified down below (so the normal modified -> flush -> set-moved 4276 * sequence applies). 4277 * 4278 * The insertion shouldn't race as this is a completely new block 4279 * and the parent is locked. 4280 */ 4281 base = NULL; /* safety, parent modify may change address */ 4282 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0); 4283 KKASSERT(parent->core.live_count < count); 4284 hammer2_chain_insert(parent, ichain, 4285 HAMMER2_CHAIN_INSERT_SPIN | 4286 HAMMER2_CHAIN_INSERT_LIVE, 4287 0); 4288 4289 /* 4290 * Make sure flushes propogate after our manual insertion. 4291 */ 4292 hammer2_chain_setflush(ichain); 4293 hammer2_chain_setflush(parent); 4294 4295 /* 4296 * Figure out what to return. 4297 */ 4298 if (rounddown2(create_key ^ key, (hammer2_key_t)1 << keybits)) { 4299 /* 4300 * Key being created is outside the key range, 4301 * return the original parent. 4302 */ 4303 hammer2_chain_unlock(ichain); 4304 hammer2_chain_drop(ichain); 4305 } else { 4306 /* 4307 * Otherwise its in the range, return the new parent. 4308 * (leave both the new and old parent locked). 4309 */ 4310 parent = ichain; 4311 } 4312 4313 return(parent); 4314 } 4315 4316 /* 4317 * Do maintenance on an indirect chain. Both parent and chain are locked. 4318 * 4319 * Returns non-zero if (chain) is deleted, either due to being empty or 4320 * because its children were safely moved into the parent. 4321 */ 4322 int 4323 hammer2_chain_indirect_maintenance(hammer2_chain_t *parent, 4324 hammer2_chain_t *chain) 4325 { 4326 hammer2_blockref_t *chain_base; 4327 hammer2_blockref_t *base; 4328 hammer2_blockref_t *bref; 4329 hammer2_blockref_t bsave; 4330 hammer2_key_t key_next; 4331 hammer2_key_t key_beg; 4332 hammer2_key_t key_end; 4333 hammer2_chain_t *sub; 4334 int chain_count; 4335 int count; 4336 int error; 4337 int generation; 4338 4339 /* 4340 * Make sure we have an accurate live_count 4341 */ 4342 if ((chain->flags & (HAMMER2_CHAIN_INITIAL | 4343 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) { 4344 base = &chain->data->npdata[0]; 4345 count = chain->bytes / sizeof(hammer2_blockref_t); 4346 hammer2_chain_countbrefs(chain, base, count); 4347 } 4348 4349 /* 4350 * If the indirect block is empty we can delete it. 4351 * (ignore deletion error) 4352 */ 4353 if (chain->core.live_count == 0 && RB_EMPTY(&chain->core.rbtree)) { 4354 hammer2_chain_delete(parent, chain, 4355 chain->bref.modify_tid, 4356 HAMMER2_DELETE_PERMANENT); 4357 hammer2_chain_repchange(parent, chain); 4358 return 1; 4359 } 4360 4361 base = hammer2_chain_base_and_count(parent, &count); 4362 4363 if ((parent->flags & (HAMMER2_CHAIN_INITIAL | 4364 HAMMER2_CHAIN_COUNTEDBREFS)) == 0) { 4365 hammer2_chain_countbrefs(parent, base, count); 4366 } 4367 4368 /* 4369 * Determine if we can collapse chain into parent, calculate 4370 * hysteresis for chain emptiness. 4371 */ 4372 if (parent->core.live_count + chain->core.live_count - 1 > count) 4373 return 0; 4374 chain_count = chain->bytes / sizeof(hammer2_blockref_t); 4375 if (chain->core.live_count > chain_count * 3 / 4) 4376 return 0; 4377 4378 /* 4379 * Ok, theoretically we can collapse chain's contents into 4380 * parent. chain is locked, but any in-memory children of chain 4381 * are not. For this to work, we must be able to dispose of any 4382 * in-memory children of chain. 4383 * 4384 * For now require that there are no in-memory children of chain. 4385 * 4386 * WARNING! Both chain and parent must remain locked across this 4387 * entire operation. 4388 */ 4389 4390 /* 4391 * Parent must be marked modified. Don't try to collapse it if we 4392 * can't mark it modified. Once modified, destroy chain to make room 4393 * and to get rid of what will be a conflicting key (this is included 4394 * in the calculation above). Finally, move the children of chain 4395 * into chain's parent. 4396 * 4397 * This order creates an accounting problem for bref.embed.stats 4398 * because we destroy chain before we remove its children. Any 4399 * elements whos blockref is already synchronized will be counted 4400 * twice. To deal with the problem we clean out chain's stats prior 4401 * to deleting it. 4402 */ 4403 error = hammer2_chain_modify(parent, 0, 0, 0); 4404 if (error) { 4405 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n", 4406 hammer2_error_str(error)); 4407 return 0; 4408 } 4409 error = hammer2_chain_modify(chain, chain->bref.modify_tid, 0, 0); 4410 if (error) { 4411 krateprintf(&krate_h2me, "hammer2: indirect_maint: %s\n", 4412 hammer2_error_str(error)); 4413 return 0; 4414 } 4415 4416 chain->bref.embed.stats.inode_count = 0; 4417 chain->bref.embed.stats.data_count = 0; 4418 error = hammer2_chain_delete(parent, chain, 4419 chain->bref.modify_tid, 4420 HAMMER2_DELETE_PERMANENT); 4421 KKASSERT(error == 0); 4422 4423 /* 4424 * The combined_find call requires core.spin to be held. One would 4425 * think there wouldn't be any conflicts since we hold chain 4426 * exclusively locked, but the caching mechanism for 0-ref children 4427 * does not require a chain lock. 4428 */ 4429 hammer2_spin_ex(&chain->core.spin); 4430 4431 key_next = 0; 4432 key_beg = 0; 4433 key_end = HAMMER2_KEY_MAX; 4434 for (;;) { 4435 chain_base = &chain->data->npdata[0]; 4436 chain_count = chain->bytes / sizeof(hammer2_blockref_t); 4437 sub = hammer2_combined_find(chain, chain_base, chain_count, 4438 &key_next, 4439 key_beg, key_end, 4440 &bref); 4441 generation = chain->core.generation; 4442 if (bref == NULL) 4443 break; 4444 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4445 4446 bsave = *bref; 4447 if (sub) { 4448 hammer2_chain_ref(sub); 4449 hammer2_spin_unex(&chain->core.spin); 4450 hammer2_chain_lock(sub, HAMMER2_RESOLVE_NEVER); 4451 } else { 4452 hammer2_spin_unex(&chain->core.spin); 4453 sub = hammer2_chain_get(chain, generation, &bsave, 4454 HAMMER2_RESOLVE_NEVER); 4455 if (sub == NULL) { 4456 hammer2_spin_ex(&chain->core.spin); 4457 continue; 4458 } 4459 } 4460 if (bcmp(&bsave, &sub->bref, sizeof(bsave)) || 4461 sub->parent != chain || 4462 (sub->flags & HAMMER2_CHAIN_DELETED)) { 4463 hammer2_chain_unlock(sub); 4464 hammer2_chain_drop(sub); 4465 hammer2_spin_ex(&chain->core.spin); 4466 sub = NULL; /* safety */ 4467 continue; 4468 } 4469 error = hammer2_chain_delete_obref(chain, sub, 4470 sub->bref.modify_tid, 0, 4471 &bsave); 4472 KKASSERT(error == 0); 4473 hammer2_chain_rename_obref(&parent, sub, 4474 sub->bref.modify_tid, 4475 HAMMER2_INSERT_SAMEPARENT, &bsave); 4476 hammer2_chain_unlock(sub); 4477 hammer2_chain_drop(sub); 4478 hammer2_spin_ex(&chain->core.spin); 4479 4480 if (key_next == 0) 4481 break; 4482 key_beg = key_next; 4483 } 4484 hammer2_spin_unex(&chain->core.spin); 4485 4486 hammer2_chain_repchange(parent, chain); 4487 4488 return 1; 4489 } 4490 4491 /* 4492 * Freemap indirect blocks 4493 * 4494 * Calculate the keybits and highside/lowside of the freemap node the 4495 * caller is creating. 4496 * 4497 * This routine will specify the next higher-level freemap key/radix 4498 * representing the lowest-ordered set. By doing so, eventually all 4499 * low-ordered sets will be moved one level down. 4500 * 4501 * We have to be careful here because the freemap reserves a limited 4502 * number of blocks for a limited number of levels. So we can't just 4503 * push indiscriminately. 4504 */ 4505 int 4506 hammer2_chain_indkey_freemap(hammer2_chain_t *parent, hammer2_key_t *keyp, 4507 int keybits, hammer2_blockref_t *base, int count) 4508 { 4509 hammer2_chain_t *chain; 4510 hammer2_blockref_t *bref; 4511 hammer2_key_t key; 4512 hammer2_key_t key_beg; 4513 hammer2_key_t key_end; 4514 hammer2_key_t key_next; 4515 int locount; 4516 int hicount; 4517 int maxloops = 300000; 4518 4519 key = *keyp; 4520 locount = 0; 4521 hicount = 0; 4522 keybits = 64; 4523 4524 /* 4525 * Calculate the range of keys in the array being careful to skip 4526 * slots which are overridden with a deletion. 4527 */ 4528 key_beg = 0; 4529 key_end = HAMMER2_KEY_MAX; 4530 hammer2_spin_ex(&parent->core.spin); 4531 4532 for (;;) { 4533 if (--maxloops == 0) { 4534 panic("indkey_freemap shit %p %p:%d\n", 4535 parent, base, count); 4536 } 4537 chain = hammer2_combined_find(parent, base, count, 4538 &key_next, 4539 key_beg, key_end, 4540 &bref); 4541 4542 /* 4543 * Exhausted search 4544 */ 4545 if (bref == NULL) 4546 break; 4547 4548 /* 4549 * Skip deleted chains. 4550 */ 4551 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) { 4552 if (key_next == 0 || key_next > key_end) 4553 break; 4554 key_beg = key_next; 4555 continue; 4556 } 4557 4558 /* 4559 * Use the full live (not deleted) element for the scan 4560 * iteration. HAMMER2 does not allow partial replacements. 4561 * 4562 * XXX should be built into hammer2_combined_find(). 4563 */ 4564 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4565 4566 if (keybits > bref->keybits) { 4567 key = bref->key; 4568 keybits = bref->keybits; 4569 } else if (keybits == bref->keybits && bref->key < key) { 4570 key = bref->key; 4571 } 4572 if (key_next == 0) 4573 break; 4574 key_beg = key_next; 4575 } 4576 hammer2_spin_unex(&parent->core.spin); 4577 4578 /* 4579 * Return the keybits for a higher-level FREEMAP_NODE covering 4580 * this node. 4581 */ 4582 switch(keybits) { 4583 case HAMMER2_FREEMAP_LEVEL0_RADIX: 4584 keybits = HAMMER2_FREEMAP_LEVEL1_RADIX; 4585 break; 4586 case HAMMER2_FREEMAP_LEVEL1_RADIX: 4587 keybits = HAMMER2_FREEMAP_LEVEL2_RADIX; 4588 break; 4589 case HAMMER2_FREEMAP_LEVEL2_RADIX: 4590 keybits = HAMMER2_FREEMAP_LEVEL3_RADIX; 4591 break; 4592 case HAMMER2_FREEMAP_LEVEL3_RADIX: 4593 keybits = HAMMER2_FREEMAP_LEVEL4_RADIX; 4594 break; 4595 case HAMMER2_FREEMAP_LEVEL4_RADIX: 4596 keybits = HAMMER2_FREEMAP_LEVEL5_RADIX; 4597 break; 4598 case HAMMER2_FREEMAP_LEVEL5_RADIX: 4599 panic("hammer2_chain_indkey_freemap: level too high"); 4600 break; 4601 default: 4602 panic("hammer2_chain_indkey_freemap: bad radix"); 4603 break; 4604 } 4605 *keyp = key; 4606 4607 return (keybits); 4608 } 4609 4610 /* 4611 * File indirect blocks 4612 * 4613 * Calculate the key/keybits for the indirect block to create by scanning 4614 * existing keys. The key being created is also passed in *keyp and can be 4615 * inside or outside the indirect block. Regardless, the indirect block 4616 * must hold at least two keys in order to guarantee sufficient space. 4617 * 4618 * We use a modified version of the freemap's fixed radix tree, but taylored 4619 * for file data. Basically we configure an indirect block encompassing the 4620 * smallest key. 4621 */ 4622 static int 4623 hammer2_chain_indkey_file(hammer2_chain_t *parent, hammer2_key_t *keyp, 4624 int keybits, hammer2_blockref_t *base, int count, 4625 int ncount) 4626 { 4627 hammer2_chain_t *chain; 4628 hammer2_blockref_t *bref; 4629 hammer2_key_t key; 4630 hammer2_key_t key_beg; 4631 hammer2_key_t key_end; 4632 hammer2_key_t key_next; 4633 int nradix; 4634 int locount; 4635 int hicount; 4636 int maxloops = 300000; 4637 4638 key = *keyp; 4639 locount = 0; 4640 hicount = 0; 4641 keybits = 64; 4642 4643 /* 4644 * Calculate the range of keys in the array being careful to skip 4645 * slots which are overridden with a deletion. 4646 * 4647 * Locate the smallest key. 4648 */ 4649 key_beg = 0; 4650 key_end = HAMMER2_KEY_MAX; 4651 hammer2_spin_ex(&parent->core.spin); 4652 4653 for (;;) { 4654 if (--maxloops == 0) { 4655 panic("indkey_freemap shit %p %p:%d\n", 4656 parent, base, count); 4657 } 4658 chain = hammer2_combined_find(parent, base, count, 4659 &key_next, 4660 key_beg, key_end, 4661 &bref); 4662 4663 /* 4664 * Exhausted search 4665 */ 4666 if (bref == NULL) 4667 break; 4668 4669 /* 4670 * Skip deleted chains. 4671 */ 4672 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) { 4673 if (key_next == 0 || key_next > key_end) 4674 break; 4675 key_beg = key_next; 4676 continue; 4677 } 4678 4679 /* 4680 * Use the full live (not deleted) element for the scan 4681 * iteration. HAMMER2 does not allow partial replacements. 4682 * 4683 * XXX should be built into hammer2_combined_find(). 4684 */ 4685 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4686 4687 if (keybits > bref->keybits) { 4688 key = bref->key; 4689 keybits = bref->keybits; 4690 } else if (keybits == bref->keybits && bref->key < key) { 4691 key = bref->key; 4692 } 4693 if (key_next == 0) 4694 break; 4695 key_beg = key_next; 4696 } 4697 hammer2_spin_unex(&parent->core.spin); 4698 4699 /* 4700 * Calculate the static keybits for a higher-level indirect block 4701 * that contains the key. 4702 */ 4703 *keyp = key; 4704 4705 switch(ncount) { 4706 case HAMMER2_IND_COUNT_MIN: 4707 nradix = HAMMER2_IND_RADIX_MIN - HAMMER2_BLOCKREF_RADIX; 4708 break; 4709 case HAMMER2_IND_COUNT_NOM: 4710 nradix = HAMMER2_IND_RADIX_NOM - HAMMER2_BLOCKREF_RADIX; 4711 break; 4712 case HAMMER2_IND_COUNT_MAX: 4713 nradix = HAMMER2_IND_RADIX_MAX - HAMMER2_BLOCKREF_RADIX; 4714 break; 4715 default: 4716 panic("bad ncount %d\n", ncount); 4717 nradix = 0; 4718 break; 4719 } 4720 4721 /* 4722 * The largest radix that can be returned for an indirect block is 4723 * 63 bits. (The largest practical indirect block radix is actually 4724 * 62 bits because the top-level inode or volume root contains four 4725 * entries, but allow 63 to be returned). 4726 */ 4727 if (nradix >= 64) 4728 nradix = 63; 4729 4730 return keybits + nradix; 4731 } 4732 4733 #if 1 4734 4735 /* 4736 * Directory indirect blocks. 4737 * 4738 * Covers both the inode index (directory of inodes), and directory contents 4739 * (filenames hardlinked to inodes). 4740 * 4741 * Because directory keys are hashed we generally try to cut the space in 4742 * half. We accomodate the inode index (which tends to have linearly 4743 * increasing inode numbers) by ensuring that the keyspace is at least large 4744 * enough to fill up the indirect block being created. 4745 */ 4746 static int 4747 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp, 4748 int keybits, hammer2_blockref_t *base, int count, 4749 int ncount) 4750 { 4751 hammer2_blockref_t *bref; 4752 hammer2_chain_t *chain; 4753 hammer2_key_t key_beg; 4754 hammer2_key_t key_end; 4755 hammer2_key_t key_next; 4756 hammer2_key_t key; 4757 int nkeybits; 4758 int locount; 4759 int hicount; 4760 int maxloops = 300000; 4761 4762 /* 4763 * NOTE: We can't take a shortcut here anymore for inodes because 4764 * the root directory can contain a mix of inodes and directory 4765 * entries (we used to just return 63 if parent->bref.type was 4766 * HAMMER2_BREF_TYPE_INODE. 4767 */ 4768 key = *keyp; 4769 locount = 0; 4770 hicount = 0; 4771 4772 /* 4773 * Calculate the range of keys in the array being careful to skip 4774 * slots which are overridden with a deletion. 4775 */ 4776 key_beg = 0; 4777 key_end = HAMMER2_KEY_MAX; 4778 hammer2_spin_ex(&parent->core.spin); 4779 4780 for (;;) { 4781 if (--maxloops == 0) { 4782 panic("indkey_freemap shit %p %p:%d\n", 4783 parent, base, count); 4784 } 4785 chain = hammer2_combined_find(parent, base, count, 4786 &key_next, 4787 key_beg, key_end, 4788 &bref); 4789 4790 /* 4791 * Exhausted search 4792 */ 4793 if (bref == NULL) 4794 break; 4795 4796 /* 4797 * Deleted object 4798 */ 4799 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) { 4800 if (key_next == 0 || key_next > key_end) 4801 break; 4802 key_beg = key_next; 4803 continue; 4804 } 4805 4806 /* 4807 * Use the full live (not deleted) element for the scan 4808 * iteration. HAMMER2 does not allow partial replacements. 4809 * 4810 * XXX should be built into hammer2_combined_find(). 4811 */ 4812 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4813 4814 /* 4815 * Expand our calculated key range (key, keybits) to fit 4816 * the scanned key. nkeybits represents the full range 4817 * that we will later cut in half (two halves @ nkeybits - 1). 4818 */ 4819 nkeybits = keybits; 4820 if (nkeybits < bref->keybits) { 4821 if (bref->keybits > 64) { 4822 kprintf("bad bref chain %p bref %p\n", 4823 chain, bref); 4824 Debugger("fubar"); 4825 } 4826 nkeybits = bref->keybits; 4827 } 4828 while (nkeybits < 64 && 4829 rounddown2(key ^ bref->key, (hammer2_key_t)1 << nkeybits) != 0) { 4830 ++nkeybits; 4831 } 4832 4833 /* 4834 * If the new key range is larger we have to determine 4835 * which side of the new key range the existing keys fall 4836 * under by checking the high bit, then collapsing the 4837 * locount into the hicount or vise-versa. 4838 */ 4839 if (keybits != nkeybits) { 4840 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) { 4841 hicount += locount; 4842 locount = 0; 4843 } else { 4844 locount += hicount; 4845 hicount = 0; 4846 } 4847 keybits = nkeybits; 4848 } 4849 4850 /* 4851 * The newly scanned key will be in the lower half or the 4852 * upper half of the (new) key range. 4853 */ 4854 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key) 4855 ++hicount; 4856 else 4857 ++locount; 4858 4859 if (key_next == 0) 4860 break; 4861 key_beg = key_next; 4862 } 4863 hammer2_spin_unex(&parent->core.spin); 4864 bref = NULL; /* now invalid (safety) */ 4865 4866 /* 4867 * Adjust keybits to represent half of the full range calculated 4868 * above (radix 63 max) for our new indirect block. 4869 */ 4870 --keybits; 4871 4872 /* 4873 * Expand keybits to hold at least ncount elements. ncount will be 4874 * a power of 2. This is to try to completely fill leaf nodes (at 4875 * least for keys which are not hashes). 4876 * 4877 * We aren't counting 'in' or 'out', we are counting 'high side' 4878 * and 'low side' based on the bit at (1LL << keybits). We want 4879 * everything to be inside in these cases so shift it all to 4880 * the low or high side depending on the new high bit. 4881 */ 4882 while (((hammer2_key_t)1 << keybits) < ncount) { 4883 ++keybits; 4884 if (key & ((hammer2_key_t)1 << keybits)) { 4885 hicount += locount; 4886 locount = 0; 4887 } else { 4888 locount += hicount; 4889 hicount = 0; 4890 } 4891 } 4892 4893 if (hicount > locount) 4894 key |= (hammer2_key_t)1 << keybits; 4895 else 4896 key &= ~(hammer2_key_t)1 << keybits; 4897 4898 *keyp = key; 4899 4900 return (keybits); 4901 } 4902 4903 #else 4904 4905 /* 4906 * Directory indirect blocks. 4907 * 4908 * Covers both the inode index (directory of inodes), and directory contents 4909 * (filenames hardlinked to inodes). 4910 * 4911 * Because directory keys are hashed we generally try to cut the space in 4912 * half. We accomodate the inode index (which tends to have linearly 4913 * increasing inode numbers) by ensuring that the keyspace is at least large 4914 * enough to fill up the indirect block being created. 4915 */ 4916 static int 4917 hammer2_chain_indkey_dir(hammer2_chain_t *parent, hammer2_key_t *keyp, 4918 int keybits, hammer2_blockref_t *base, int count, 4919 int ncount) 4920 { 4921 hammer2_blockref_t *bref; 4922 hammer2_chain_t *chain; 4923 hammer2_key_t key_beg; 4924 hammer2_key_t key_end; 4925 hammer2_key_t key_next; 4926 hammer2_key_t key; 4927 int nkeybits; 4928 int locount; 4929 int hicount; 4930 int maxloops = 300000; 4931 4932 /* 4933 * Shortcut if the parent is the inode. In this situation the 4934 * parent has 4+1 directory entries and we are creating an indirect 4935 * block capable of holding many more. 4936 */ 4937 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) { 4938 return 63; 4939 } 4940 4941 key = *keyp; 4942 locount = 0; 4943 hicount = 0; 4944 4945 /* 4946 * Calculate the range of keys in the array being careful to skip 4947 * slots which are overridden with a deletion. 4948 */ 4949 key_beg = 0; 4950 key_end = HAMMER2_KEY_MAX; 4951 hammer2_spin_ex(&parent->core.spin); 4952 4953 for (;;) { 4954 if (--maxloops == 0) { 4955 panic("indkey_freemap shit %p %p:%d\n", 4956 parent, base, count); 4957 } 4958 chain = hammer2_combined_find(parent, base, count, 4959 &key_next, 4960 key_beg, key_end, 4961 &bref); 4962 4963 /* 4964 * Exhausted search 4965 */ 4966 if (bref == NULL) 4967 break; 4968 4969 /* 4970 * Deleted object 4971 */ 4972 if (chain && (chain->flags & HAMMER2_CHAIN_DELETED)) { 4973 if (key_next == 0 || key_next > key_end) 4974 break; 4975 key_beg = key_next; 4976 continue; 4977 } 4978 4979 /* 4980 * Use the full live (not deleted) element for the scan 4981 * iteration. HAMMER2 does not allow partial replacements. 4982 * 4983 * XXX should be built into hammer2_combined_find(). 4984 */ 4985 key_next = bref->key + ((hammer2_key_t)1 << bref->keybits); 4986 4987 /* 4988 * Expand our calculated key range (key, keybits) to fit 4989 * the scanned key. nkeybits represents the full range 4990 * that we will later cut in half (two halves @ nkeybits - 1). 4991 */ 4992 nkeybits = keybits; 4993 if (nkeybits < bref->keybits) { 4994 if (bref->keybits > 64) { 4995 kprintf("bad bref chain %p bref %p\n", 4996 chain, bref); 4997 Debugger("fubar"); 4998 } 4999 nkeybits = bref->keybits; 5000 } 5001 while (nkeybits < 64 && 5002 (~(((hammer2_key_t)1 << nkeybits) - 1) & 5003 (key ^ bref->key)) != 0) { 5004 ++nkeybits; 5005 } 5006 5007 /* 5008 * If the new key range is larger we have to determine 5009 * which side of the new key range the existing keys fall 5010 * under by checking the high bit, then collapsing the 5011 * locount into the hicount or vise-versa. 5012 */ 5013 if (keybits != nkeybits) { 5014 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) { 5015 hicount += locount; 5016 locount = 0; 5017 } else { 5018 locount += hicount; 5019 hicount = 0; 5020 } 5021 keybits = nkeybits; 5022 } 5023 5024 /* 5025 * The newly scanned key will be in the lower half or the 5026 * upper half of the (new) key range. 5027 */ 5028 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key) 5029 ++hicount; 5030 else 5031 ++locount; 5032 5033 if (key_next == 0) 5034 break; 5035 key_beg = key_next; 5036 } 5037 hammer2_spin_unex(&parent->core.spin); 5038 bref = NULL; /* now invalid (safety) */ 5039 5040 /* 5041 * Adjust keybits to represent half of the full range calculated 5042 * above (radix 63 max) for our new indirect block. 5043 */ 5044 --keybits; 5045 5046 /* 5047 * Expand keybits to hold at least ncount elements. ncount will be 5048 * a power of 2. This is to try to completely fill leaf nodes (at 5049 * least for keys which are not hashes). 5050 * 5051 * We aren't counting 'in' or 'out', we are counting 'high side' 5052 * and 'low side' based on the bit at (1LL << keybits). We want 5053 * everything to be inside in these cases so shift it all to 5054 * the low or high side depending on the new high bit. 5055 */ 5056 while (((hammer2_key_t)1 << keybits) < ncount) { 5057 ++keybits; 5058 if (key & ((hammer2_key_t)1 << keybits)) { 5059 hicount += locount; 5060 locount = 0; 5061 } else { 5062 locount += hicount; 5063 hicount = 0; 5064 } 5065 } 5066 5067 if (hicount > locount) 5068 key |= (hammer2_key_t)1 << keybits; 5069 else 5070 key &= ~(hammer2_key_t)1 << keybits; 5071 5072 *keyp = key; 5073 5074 return (keybits); 5075 } 5076 5077 #endif 5078 5079 /* 5080 * Sets CHAIN_DELETED and remove the chain's blockref from the parent if 5081 * it exists. 5082 * 5083 * Both parent and chain must be locked exclusively. 5084 * 5085 * This function will modify the parent if the blockref requires removal 5086 * from the parent's block table. 5087 * 5088 * This function is NOT recursive. Any entity already pushed into the 5089 * chain (such as an inode) may still need visibility into its contents, 5090 * as well as the ability to read and modify the contents. For example, 5091 * for an unlinked file which is still open. 5092 * 5093 * Also note that the flusher is responsible for cleaning up empty 5094 * indirect blocks. 5095 */ 5096 int 5097 hammer2_chain_delete(hammer2_chain_t *parent, hammer2_chain_t *chain, 5098 hammer2_tid_t mtid, int flags) 5099 { 5100 int error = 0; 5101 5102 KKASSERT(hammer2_mtx_owned(&chain->lock)); 5103 5104 /* 5105 * Nothing to do if already marked. 5106 * 5107 * We need the spinlock on the core whos RBTREE contains chain 5108 * to protect against races. 5109 */ 5110 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) { 5111 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 && 5112 chain->parent == parent); 5113 error = _hammer2_chain_delete_helper(parent, chain, 5114 mtid, flags, NULL); 5115 } 5116 5117 /* 5118 * Permanent deletions mark the chain as destroyed. 5119 * 5120 * NOTE: We do not setflush the chain unless the deletion is 5121 * permanent, since the deletion of a chain does not actually 5122 * require it to be flushed. 5123 */ 5124 if (error == 0) { 5125 if (flags & HAMMER2_DELETE_PERMANENT) { 5126 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY); 5127 hammer2_chain_setflush(chain); 5128 } 5129 } 5130 5131 return error; 5132 } 5133 5134 static int 5135 hammer2_chain_delete_obref(hammer2_chain_t *parent, hammer2_chain_t *chain, 5136 hammer2_tid_t mtid, int flags, 5137 hammer2_blockref_t *obref) 5138 { 5139 int error = 0; 5140 5141 KKASSERT(hammer2_mtx_owned(&chain->lock)); 5142 5143 /* 5144 * Nothing to do if already marked. 5145 * 5146 * We need the spinlock on the core whos RBTREE contains chain 5147 * to protect against races. 5148 */ 5149 obref->type = HAMMER2_BREF_TYPE_EMPTY; 5150 if ((chain->flags & HAMMER2_CHAIN_DELETED) == 0) { 5151 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0 && 5152 chain->parent == parent); 5153 error = _hammer2_chain_delete_helper(parent, chain, 5154 mtid, flags, obref); 5155 } 5156 5157 /* 5158 * Permanent deletions mark the chain as destroyed. 5159 * 5160 * NOTE: We do not setflush the chain unless the deletion is 5161 * permanent, since the deletion of a chain does not actually 5162 * require it to be flushed. 5163 */ 5164 if (error == 0) { 5165 if (flags & HAMMER2_DELETE_PERMANENT) { 5166 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DESTROY); 5167 hammer2_chain_setflush(chain); 5168 } 5169 } 5170 5171 return error; 5172 } 5173 5174 /* 5175 * Returns the index of the nearest element in the blockref array >= elm. 5176 * Returns (count) if no element could be found. 5177 * 5178 * Sets *key_nextp to the next key for loop purposes but does not modify 5179 * it if the next key would be higher than the current value of *key_nextp. 5180 * Note that *key_nexp can overflow to 0, which should be tested by the 5181 * caller. 5182 * 5183 * WARNING! Must be called with parent's spinlock held. Spinlock remains 5184 * held through the operation. 5185 */ 5186 static int 5187 hammer2_base_find(hammer2_chain_t *parent, 5188 hammer2_blockref_t *base, int count, 5189 hammer2_key_t *key_nextp, 5190 hammer2_key_t key_beg, hammer2_key_t key_end) 5191 { 5192 hammer2_blockref_t *scan; 5193 hammer2_key_t scan_end; 5194 int i; 5195 int limit; 5196 5197 /* 5198 * Require the live chain's already have their core's counted 5199 * so we can optimize operations. 5200 */ 5201 KKASSERT(parent->flags & HAMMER2_CHAIN_COUNTEDBREFS); 5202 5203 /* 5204 * Degenerate case 5205 */ 5206 if (count == 0 || base == NULL) 5207 return(count); 5208 5209 /* 5210 * Sequential optimization using parent->cache_index. This is 5211 * the most likely scenario. 5212 * 5213 * We can avoid trailing empty entries on live chains, otherwise 5214 * we might have to check the whole block array. 5215 */ 5216 i = parent->cache_index; /* SMP RACE OK */ 5217 cpu_ccfence(); 5218 limit = parent->core.live_zero; 5219 if (i >= limit) 5220 i = limit - 1; 5221 if (i < 0) 5222 i = 0; 5223 KKASSERT(i < count); 5224 5225 /* 5226 * Search backwards 5227 */ 5228 scan = &base[i]; 5229 while (i > 0 && (scan->type == HAMMER2_BREF_TYPE_EMPTY || 5230 scan->key > key_beg)) { 5231 --scan; 5232 --i; 5233 } 5234 parent->cache_index = i; 5235 5236 /* 5237 * Search forwards, stop when we find a scan element which 5238 * encloses the key or until we know that there are no further 5239 * elements. 5240 */ 5241 while (i < count) { 5242 if (scan->type != HAMMER2_BREF_TYPE_EMPTY) { 5243 scan_end = scan->key + 5244 ((hammer2_key_t)1 << scan->keybits) - 1; 5245 if (scan->key > key_beg || scan_end >= key_beg) 5246 break; 5247 } 5248 if (i >= limit) 5249 return (count); 5250 ++scan; 5251 ++i; 5252 } 5253 if (i != count) { 5254 parent->cache_index = i; 5255 if (i >= limit) { 5256 i = count; 5257 } else { 5258 scan_end = scan->key + 5259 ((hammer2_key_t)1 << scan->keybits); 5260 if (scan_end && (*key_nextp > scan_end || 5261 *key_nextp == 0)) { 5262 *key_nextp = scan_end; 5263 } 5264 } 5265 } 5266 return (i); 5267 } 5268 5269 /* 5270 * Do a combined search and return the next match either from the blockref 5271 * array or from the in-memory chain. Sets *brefp to the returned bref in 5272 * both cases, or sets it to NULL if the search exhausted. Only returns 5273 * a non-NULL chain if the search matched from the in-memory chain. 5274 * 5275 * When no in-memory chain has been found and a non-NULL bref is returned 5276 * in *brefp. 5277 * 5278 * 5279 * The returned chain is not locked or referenced. Use the returned bref 5280 * to determine if the search exhausted or not. Iterate if the base find 5281 * is chosen but matches a deleted chain. 5282 * 5283 * WARNING! Must be called with parent's spinlock held. Spinlock remains 5284 * held through the operation. 5285 */ 5286 static hammer2_chain_t * 5287 hammer2_combined_find(hammer2_chain_t *parent, 5288 hammer2_blockref_t *base, int count, 5289 hammer2_key_t *key_nextp, 5290 hammer2_key_t key_beg, hammer2_key_t key_end, 5291 hammer2_blockref_t **brefp) 5292 { 5293 hammer2_blockref_t *bref; 5294 hammer2_chain_t *chain; 5295 int i; 5296 5297 /* 5298 * Lookup in block array and in rbtree. 5299 */ 5300 *key_nextp = key_end + 1; 5301 i = hammer2_base_find(parent, base, count, key_nextp, 5302 key_beg, key_end); 5303 chain = hammer2_chain_find(parent, key_nextp, key_beg, key_end); 5304 5305 /* 5306 * Neither matched 5307 */ 5308 if (i == count && chain == NULL) { 5309 *brefp = NULL; 5310 return(NULL); 5311 } 5312 5313 /* 5314 * Only chain matched. 5315 */ 5316 if (i == count) { 5317 bref = &chain->bref; 5318 goto found; 5319 } 5320 5321 /* 5322 * Only blockref matched. 5323 */ 5324 if (chain == NULL) { 5325 bref = &base[i]; 5326 goto found; 5327 } 5328 5329 /* 5330 * Both in-memory and blockref matched, select the nearer element. 5331 * 5332 * If both are flush with the left-hand side or both are the 5333 * same distance away, select the chain. In this situation the 5334 * chain must have been loaded from the matching blockmap. 5335 */ 5336 if ((chain->bref.key <= key_beg && base[i].key <= key_beg) || 5337 chain->bref.key == base[i].key) { 5338 KKASSERT(chain->bref.key == base[i].key); 5339 bref = &chain->bref; 5340 goto found; 5341 } 5342 5343 /* 5344 * Select the nearer key 5345 */ 5346 if (chain->bref.key < base[i].key) { 5347 bref = &chain->bref; 5348 } else { 5349 bref = &base[i]; 5350 chain = NULL; 5351 } 5352 5353 /* 5354 * If the bref is out of bounds we've exhausted our search. 5355 */ 5356 found: 5357 if (bref->key > key_end) { 5358 *brefp = NULL; 5359 chain = NULL; 5360 } else { 5361 *brefp = bref; 5362 } 5363 return(chain); 5364 } 5365 5366 /* 5367 * Locate the specified block array element and delete it. The element 5368 * must exist. 5369 * 5370 * The spin lock on the related chain must be held. 5371 * 5372 * NOTE: live_count was adjusted when the chain was deleted, so it does not 5373 * need to be adjusted when we commit the media change. 5374 */ 5375 void 5376 hammer2_base_delete(hammer2_chain_t *parent, 5377 hammer2_blockref_t *base, int count, 5378 hammer2_chain_t *chain, 5379 hammer2_blockref_t *obref) 5380 { 5381 hammer2_blockref_t *elm = &chain->bref; 5382 hammer2_blockref_t *scan; 5383 hammer2_key_t key_next; 5384 int i; 5385 5386 /* 5387 * Delete element. Expect the element to exist. 5388 * 5389 * XXX see caller, flush code not yet sophisticated enough to prevent 5390 * re-flushed in some cases. 5391 */ 5392 key_next = 0; /* max range */ 5393 i = hammer2_base_find(parent, base, count, &key_next, 5394 elm->key, elm->key); 5395 scan = &base[i]; 5396 if (i == count || scan->type == HAMMER2_BREF_TYPE_EMPTY || 5397 scan->key != elm->key || 5398 ((chain->flags & HAMMER2_CHAIN_BLKMAPUPD) == 0 && 5399 scan->keybits != elm->keybits)) { 5400 hammer2_spin_unex(&parent->core.spin); 5401 panic("delete base %p element not found at %d/%d elm %p\n", 5402 base, i, count, elm); 5403 return; 5404 } 5405 5406 /* 5407 * Update stats and zero the entry. 5408 * 5409 * NOTE: Handle radix == 0 (0 bytes) case. 5410 */ 5411 if ((int)(scan->data_off & HAMMER2_OFF_MASK_RADIX)) { 5412 parent->bref.embed.stats.data_count -= (hammer2_off_t)1 << 5413 (int)(scan->data_off & HAMMER2_OFF_MASK_RADIX); 5414 } 5415 switch(scan->type) { 5416 case HAMMER2_BREF_TYPE_INODE: 5417 --parent->bref.embed.stats.inode_count; 5418 /* fall through */ 5419 case HAMMER2_BREF_TYPE_DATA: 5420 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) { 5421 atomic_set_int(&chain->flags, 5422 HAMMER2_CHAIN_HINT_LEAF_COUNT); 5423 } else { 5424 if (parent->bref.leaf_count) 5425 --parent->bref.leaf_count; 5426 } 5427 /* fall through */ 5428 case HAMMER2_BREF_TYPE_INDIRECT: 5429 if (scan->type != HAMMER2_BREF_TYPE_DATA) { 5430 parent->bref.embed.stats.data_count -= 5431 scan->embed.stats.data_count; 5432 parent->bref.embed.stats.inode_count -= 5433 scan->embed.stats.inode_count; 5434 } 5435 if (scan->type == HAMMER2_BREF_TYPE_INODE) 5436 break; 5437 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) { 5438 atomic_set_int(&chain->flags, 5439 HAMMER2_CHAIN_HINT_LEAF_COUNT); 5440 } else { 5441 if (parent->bref.leaf_count <= scan->leaf_count) 5442 parent->bref.leaf_count = 0; 5443 else 5444 parent->bref.leaf_count -= scan->leaf_count; 5445 } 5446 break; 5447 case HAMMER2_BREF_TYPE_DIRENT: 5448 if (parent->bref.leaf_count == HAMMER2_BLOCKREF_LEAF_MAX) { 5449 atomic_set_int(&chain->flags, 5450 HAMMER2_CHAIN_HINT_LEAF_COUNT); 5451 } else { 5452 if (parent->bref.leaf_count) 5453 --parent->bref.leaf_count; 5454 } 5455 default: 5456 break; 5457 } 5458 5459 if (obref) 5460 *obref = *scan; 5461 bzero(scan, sizeof(*scan)); 5462 5463 /* 5464 * We can only optimize parent->core.live_zero for live chains. 5465 */ 5466 if (parent->core.live_zero == i + 1) { 5467 while (--i >= 0 && base[i].type == HAMMER2_BREF_TYPE_EMPTY) 5468 ; 5469 parent->core.live_zero = i + 1; 5470 } 5471 5472 /* 5473 * Clear appropriate blockmap flags in chain. 5474 */ 5475 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED | 5476 HAMMER2_CHAIN_BLKMAPUPD); 5477 } 5478 5479 /* 5480 * Insert the specified element. The block array must not already have the 5481 * element and must have space available for the insertion. 5482 * 5483 * The spin lock on the related chain must be held. 5484 * 5485 * NOTE: live_count was adjusted when the chain was deleted, so it does not 5486 * need to be adjusted when we commit the media change. 5487 */ 5488 void 5489 hammer2_base_insert(hammer2_chain_t *parent, 5490 hammer2_blockref_t *base, int count, 5491 hammer2_chain_t *chain, hammer2_blockref_t *elm) 5492 { 5493 hammer2_key_t key_next; 5494 hammer2_key_t xkey; 5495 int i; 5496 int j; 5497 int k; 5498 int l; 5499 int u = 1; 5500 5501 /* 5502 * Insert new element. Expect the element to not already exist 5503 * unless we are replacing it. 5504 * 5505 * XXX see caller, flush code not yet sophisticated enough to prevent 5506 * re-flushed in some cases. 5507 */ 5508 key_next = 0; /* max range */ 5509 i = hammer2_base_find(parent, base, count, &key_next, 5510 elm->key, elm->key); 5511 5512 /* 5513 * Shortcut fill optimization, typical ordered insertion(s) may not 5514 * require a search. 5515 */ 5516 KKASSERT(i >= 0 && i <= count); 5517 5518 /* 5519 * Set appropriate blockmap flags in chain (if not NULL) 5520 */ 5521 if (chain) 5522 atomic_set_int(&chain->flags, HAMMER2_CHAIN_BLKMAPPED); 5523 5524 /* 5525 * Update stats and zero the entry 5526 */ 5527 if ((int)(elm->data_off & HAMMER2_OFF_MASK_RADIX)) { 5528 parent->bref.embed.stats.data_count += (hammer2_off_t)1 << 5529 (int)(elm->data_off & HAMMER2_OFF_MASK_RADIX); 5530 } 5531 switch(elm->type) { 5532 case HAMMER2_BREF_TYPE_INODE: 5533 ++parent->bref.embed.stats.inode_count; 5534 /* fall through */ 5535 case HAMMER2_BREF_TYPE_DATA: 5536 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX) 5537 ++parent->bref.leaf_count; 5538 /* fall through */ 5539 case HAMMER2_BREF_TYPE_INDIRECT: 5540 if (elm->type != HAMMER2_BREF_TYPE_DATA) { 5541 parent->bref.embed.stats.data_count += 5542 elm->embed.stats.data_count; 5543 parent->bref.embed.stats.inode_count += 5544 elm->embed.stats.inode_count; 5545 } 5546 if (elm->type == HAMMER2_BREF_TYPE_INODE) 5547 break; 5548 if (parent->bref.leaf_count + elm->leaf_count < 5549 HAMMER2_BLOCKREF_LEAF_MAX) { 5550 parent->bref.leaf_count += elm->leaf_count; 5551 } else { 5552 parent->bref.leaf_count = HAMMER2_BLOCKREF_LEAF_MAX; 5553 } 5554 break; 5555 case HAMMER2_BREF_TYPE_DIRENT: 5556 if (parent->bref.leaf_count != HAMMER2_BLOCKREF_LEAF_MAX) 5557 ++parent->bref.leaf_count; 5558 break; 5559 default: 5560 break; 5561 } 5562 5563 5564 /* 5565 * We can only optimize parent->core.live_zero for live chains. 5566 */ 5567 if (i == count && parent->core.live_zero < count) { 5568 i = parent->core.live_zero++; 5569 base[i] = *elm; 5570 return; 5571 } 5572 5573 xkey = elm->key + ((hammer2_key_t)1 << elm->keybits) - 1; 5574 if (i != count && (base[i].key < elm->key || xkey >= base[i].key)) { 5575 hammer2_spin_unex(&parent->core.spin); 5576 panic("insert base %p overlapping elements at %d elm %p\n", 5577 base, i, elm); 5578 } 5579 5580 /* 5581 * Try to find an empty slot before or after. 5582 */ 5583 j = i; 5584 k = i; 5585 while (j > 0 || k < count) { 5586 --j; 5587 if (j >= 0 && base[j].type == HAMMER2_BREF_TYPE_EMPTY) { 5588 if (j == i - 1) { 5589 base[j] = *elm; 5590 } else { 5591 bcopy(&base[j+1], &base[j], 5592 (i - j - 1) * sizeof(*base)); 5593 base[i - 1] = *elm; 5594 } 5595 goto validate; 5596 } 5597 ++k; 5598 if (k < count && base[k].type == HAMMER2_BREF_TYPE_EMPTY) { 5599 bcopy(&base[i], &base[i+1], 5600 (k - i) * sizeof(hammer2_blockref_t)); 5601 base[i] = *elm; 5602 5603 /* 5604 * We can only update parent->core.live_zero for live 5605 * chains. 5606 */ 5607 if (parent->core.live_zero <= k) 5608 parent->core.live_zero = k + 1; 5609 u = 2; 5610 goto validate; 5611 } 5612 } 5613 panic("hammer2_base_insert: no room!"); 5614 5615 /* 5616 * Debugging 5617 */ 5618 validate: 5619 key_next = 0; 5620 for (l = 0; l < count; ++l) { 5621 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) { 5622 key_next = base[l].key + 5623 ((hammer2_key_t)1 << base[l].keybits) - 1; 5624 break; 5625 } 5626 } 5627 while (++l < count) { 5628 if (base[l].type != HAMMER2_BREF_TYPE_EMPTY) { 5629 if (base[l].key <= key_next) 5630 panic("base_insert %d %d,%d,%d fail %p:%d", u, i, j, k, base, l); 5631 key_next = base[l].key + 5632 ((hammer2_key_t)1 << base[l].keybits) - 1; 5633 5634 } 5635 } 5636 5637 } 5638 5639 #if 0 5640 5641 /* 5642 * Sort the blockref array for the chain. Used by the flush code to 5643 * sort the blockref[] array. 5644 * 5645 * The chain must be exclusively locked AND spin-locked. 5646 */ 5647 typedef hammer2_blockref_t *hammer2_blockref_p; 5648 5649 static 5650 int 5651 hammer2_base_sort_callback(const void *v1, const void *v2) 5652 { 5653 hammer2_blockref_p bref1 = *(const hammer2_blockref_p *)v1; 5654 hammer2_blockref_p bref2 = *(const hammer2_blockref_p *)v2; 5655 5656 /* 5657 * Make sure empty elements are placed at the end of the array 5658 */ 5659 if (bref1->type == HAMMER2_BREF_TYPE_EMPTY) { 5660 if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) 5661 return(0); 5662 return(1); 5663 } else if (bref2->type == HAMMER2_BREF_TYPE_EMPTY) { 5664 return(-1); 5665 } 5666 5667 /* 5668 * Sort by key 5669 */ 5670 if (bref1->key < bref2->key) 5671 return(-1); 5672 if (bref1->key > bref2->key) 5673 return(1); 5674 return(0); 5675 } 5676 5677 void 5678 hammer2_base_sort(hammer2_chain_t *chain) 5679 { 5680 hammer2_blockref_t *base; 5681 int count; 5682 5683 switch(chain->bref.type) { 5684 case HAMMER2_BREF_TYPE_INODE: 5685 /* 5686 * Special shortcut for embedded data returns the inode 5687 * itself. Callers must detect this condition and access 5688 * the embedded data (the strategy code does this for us). 5689 * 5690 * This is only applicable to regular files and softlinks. 5691 */ 5692 if (chain->data->ipdata.meta.op_flags & 5693 HAMMER2_OPFLAG_DIRECTDATA) { 5694 return; 5695 } 5696 base = &chain->data->ipdata.u.blockset.blockref[0]; 5697 count = HAMMER2_SET_COUNT; 5698 break; 5699 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 5700 case HAMMER2_BREF_TYPE_INDIRECT: 5701 /* 5702 * Optimize indirect blocks in the INITIAL state to avoid 5703 * I/O. 5704 */ 5705 KKASSERT((chain->flags & HAMMER2_CHAIN_INITIAL) == 0); 5706 base = &chain->data->npdata[0]; 5707 count = chain->bytes / sizeof(hammer2_blockref_t); 5708 break; 5709 case HAMMER2_BREF_TYPE_VOLUME: 5710 base = &chain->data->voldata.sroot_blockset.blockref[0]; 5711 count = HAMMER2_SET_COUNT; 5712 break; 5713 case HAMMER2_BREF_TYPE_FREEMAP: 5714 base = &chain->data->blkset.blockref[0]; 5715 count = HAMMER2_SET_COUNT; 5716 break; 5717 default: 5718 panic("hammer2_base_sort: unrecognized " 5719 "blockref(A) type: %d", 5720 chain->bref.type); 5721 base = NULL; /* safety */ 5722 count = 0; /* safety */ 5723 break; 5724 } 5725 kqsort(base, count, sizeof(*base), hammer2_base_sort_callback); 5726 } 5727 5728 #endif 5729 5730 /* 5731 * Set the check data for a chain. This can be a heavy-weight operation 5732 * and typically only runs on-flush. For file data check data is calculated 5733 * when the logical buffers are flushed. 5734 */ 5735 void 5736 hammer2_chain_setcheck(hammer2_chain_t *chain, void *bdata) 5737 { 5738 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_NOTTESTED); 5739 5740 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 5741 case HAMMER2_CHECK_NONE: 5742 break; 5743 case HAMMER2_CHECK_DISABLED: 5744 break; 5745 case HAMMER2_CHECK_ISCSI32: 5746 chain->bref.check.iscsi32.value = 5747 hammer2_icrc32(bdata, chain->bytes); 5748 break; 5749 case HAMMER2_CHECK_XXHASH64: 5750 chain->bref.check.xxhash64.value = 5751 XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED); 5752 break; 5753 case HAMMER2_CHECK_SHA192: 5754 { 5755 SHA256_CTX hash_ctx; 5756 union { 5757 uint8_t digest[SHA256_DIGEST_LENGTH]; 5758 uint64_t digest64[SHA256_DIGEST_LENGTH/8]; 5759 } u; 5760 5761 SHA256_Init(&hash_ctx); 5762 SHA256_Update(&hash_ctx, bdata, chain->bytes); 5763 SHA256_Final(u.digest, &hash_ctx); 5764 u.digest64[2] ^= u.digest64[3]; 5765 bcopy(u.digest, 5766 chain->bref.check.sha192.data, 5767 sizeof(chain->bref.check.sha192.data)); 5768 } 5769 break; 5770 case HAMMER2_CHECK_FREEMAP: 5771 chain->bref.check.freemap.icrc32 = 5772 hammer2_icrc32(bdata, chain->bytes); 5773 break; 5774 default: 5775 kprintf("hammer2_chain_setcheck: unknown check type %02x\n", 5776 chain->bref.methods); 5777 break; 5778 } 5779 } 5780 5781 /* 5782 * Characterize a failed check code and try to trace back to the inode. 5783 */ 5784 static void 5785 hammer2_characterize_failed_chain(hammer2_chain_t *chain, uint64_t check, 5786 int bits) 5787 { 5788 hammer2_chain_t *lchain; 5789 hammer2_chain_t *ochain; 5790 int did; 5791 5792 did = krateprintf(&krate_h2chk, 5793 "chain %016jx.%02x (%s) meth=%02x CHECK FAIL " 5794 "(flags=%08x, bref/data ", 5795 chain->bref.data_off, 5796 chain->bref.type, 5797 hammer2_bref_type_str(chain->bref.type), 5798 chain->bref.methods, 5799 chain->flags); 5800 if (did == 0) 5801 return; 5802 5803 if (bits == 32) { 5804 kprintf("%08x/%08x)\n", 5805 chain->bref.check.iscsi32.value, 5806 (uint32_t)check); 5807 } else { 5808 kprintf("%016jx/%016jx)\n", 5809 chain->bref.check.xxhash64.value, 5810 check); 5811 } 5812 5813 /* 5814 * Run up the chains to try to find the governing inode so we 5815 * can report it. 5816 * 5817 * XXX This error reporting is not really MPSAFE 5818 */ 5819 ochain = chain; 5820 lchain = chain; 5821 while (chain && chain->bref.type != HAMMER2_BREF_TYPE_INODE) { 5822 lchain = chain; 5823 chain = chain->parent; 5824 } 5825 5826 if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE && 5827 ((chain->bref.flags & HAMMER2_BREF_FLAG_PFSROOT) == 0 || 5828 (lchain->bref.key & HAMMER2_DIRHASH_VISIBLE))) { 5829 kprintf(" Resides at/in inode %ld\n", 5830 chain->bref.key); 5831 } else if (chain && chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 5832 kprintf(" Resides in inode index - CRITICAL!!!\n"); 5833 } else { 5834 kprintf(" Resides in root index - CRITICAL!!!\n"); 5835 } 5836 if (ochain->hmp) { 5837 const char *pfsname = "UNKNOWN"; 5838 int i; 5839 5840 if (ochain->pmp) { 5841 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) { 5842 if (ochain->pmp->pfs_hmps[i] == ochain->hmp && 5843 ochain->pmp->pfs_names[i]) { 5844 pfsname = ochain->pmp->pfs_names[i]; 5845 break; 5846 } 5847 } 5848 } 5849 kprintf(" In pfs %s on device %s\n", 5850 pfsname, ochain->hmp->devrepname); 5851 } 5852 } 5853 5854 /* 5855 * Returns non-zero on success, 0 on failure. 5856 */ 5857 int 5858 hammer2_chain_testcheck(hammer2_chain_t *chain, void *bdata) 5859 { 5860 uint32_t check32; 5861 uint64_t check64; 5862 int r; 5863 5864 if (chain->flags & HAMMER2_CHAIN_NOTTESTED) 5865 return 1; 5866 5867 switch(HAMMER2_DEC_CHECK(chain->bref.methods)) { 5868 case HAMMER2_CHECK_NONE: 5869 r = 1; 5870 break; 5871 case HAMMER2_CHECK_DISABLED: 5872 r = 1; 5873 break; 5874 case HAMMER2_CHECK_ISCSI32: 5875 check32 = hammer2_icrc32(bdata, chain->bytes); 5876 r = (chain->bref.check.iscsi32.value == check32); 5877 if (r == 0) { 5878 hammer2_characterize_failed_chain(chain, check32, 32); 5879 } 5880 hammer2_process_icrc32 += chain->bytes; 5881 break; 5882 case HAMMER2_CHECK_XXHASH64: 5883 check64 = XXH64(bdata, chain->bytes, XXH_HAMMER2_SEED); 5884 r = (chain->bref.check.xxhash64.value == check64); 5885 if (r == 0) { 5886 hammer2_characterize_failed_chain(chain, check64, 64); 5887 } 5888 hammer2_process_xxhash64 += chain->bytes; 5889 break; 5890 case HAMMER2_CHECK_SHA192: 5891 { 5892 SHA256_CTX hash_ctx; 5893 union { 5894 uint8_t digest[SHA256_DIGEST_LENGTH]; 5895 uint64_t digest64[SHA256_DIGEST_LENGTH/8]; 5896 } u; 5897 5898 SHA256_Init(&hash_ctx); 5899 SHA256_Update(&hash_ctx, bdata, chain->bytes); 5900 SHA256_Final(u.digest, &hash_ctx); 5901 u.digest64[2] ^= u.digest64[3]; 5902 if (bcmp(u.digest, 5903 chain->bref.check.sha192.data, 5904 sizeof(chain->bref.check.sha192.data)) == 0) { 5905 r = 1; 5906 } else { 5907 r = 0; 5908 krateprintf(&krate_h2chk, 5909 "chain %016jx.%02x meth=%02x " 5910 "CHECK FAIL\n", 5911 chain->bref.data_off, 5912 chain->bref.type, 5913 chain->bref.methods); 5914 } 5915 } 5916 break; 5917 case HAMMER2_CHECK_FREEMAP: 5918 r = (chain->bref.check.freemap.icrc32 == 5919 hammer2_icrc32(bdata, chain->bytes)); 5920 if (r == 0) { 5921 int did; 5922 5923 did = krateprintf(&krate_h2chk, 5924 "chain %016jx.%02x meth=%02x " 5925 "CHECK FAIL\n", 5926 chain->bref.data_off, 5927 chain->bref.type, 5928 chain->bref.methods); 5929 if (did) { 5930 kprintf("freemap.icrc %08x icrc32 %08x (%d)\n", 5931 chain->bref.check.freemap.icrc32, 5932 hammer2_icrc32(bdata, chain->bytes), 5933 chain->bytes); 5934 if (chain->dio) { 5935 kprintf("dio %p buf %016jx,%d " 5936 "bdata %p/%p\n", 5937 chain->dio, 5938 chain->dio->bp->b_loffset, 5939 chain->dio->bp->b_bufsize, 5940 bdata, 5941 chain->dio->bp->b_data); 5942 } 5943 } 5944 } 5945 break; 5946 default: 5947 kprintf("hammer2_chain_testcheck: unknown check type %02x\n", 5948 chain->bref.methods); 5949 r = 1; 5950 break; 5951 } 5952 return r; 5953 } 5954 5955 /* 5956 * Acquire the chain and parent representing the specified inode for the 5957 * device at the specified cluster index. 5958 * 5959 * The flags passed in are LOOKUP flags, not RESOLVE flags. 5960 * 5961 * If we are unable to locate the inode, HAMMER2_ERROR_EIO or HAMMER2_ERROR_CHECK 5962 * is returned. In case of error, *chainp and/or *parentp may still be returned 5963 * non-NULL. 5964 * 5965 * The caller may pass-in a locked *parentp and/or *chainp, or neither. 5966 * They will be unlocked and released by this function. The *parentp and 5967 * *chainp representing the located inode are returned locked. 5968 * 5969 * The returned error includes any error on the returned chain in addition to 5970 * errors incurred while trying to lookup the inode. However, a chain->error 5971 * might not be recognized if HAMMER2_LOOKUP_NODATA is passed. This flag may 5972 * not be passed to this function. 5973 */ 5974 int 5975 hammer2_chain_inode_find(hammer2_pfs_t *pmp, hammer2_key_t inum, 5976 int clindex, int flags, 5977 hammer2_chain_t **parentp, hammer2_chain_t **chainp) 5978 { 5979 hammer2_chain_t *parent; 5980 hammer2_chain_t *rchain; 5981 hammer2_key_t key_dummy; 5982 hammer2_inode_t *ip; 5983 int resolve_flags; 5984 int error; 5985 5986 KKASSERT((flags & HAMMER2_LOOKUP_NODATA) == 0); 5987 5988 resolve_flags = (flags & HAMMER2_LOOKUP_SHARED) ? 5989 HAMMER2_RESOLVE_SHARED : 0; 5990 5991 /* 5992 * Caller expects us to replace these. 5993 */ 5994 if (*chainp) { 5995 hammer2_chain_unlock(*chainp); 5996 hammer2_chain_drop(*chainp); 5997 *chainp = NULL; 5998 } 5999 if (*parentp) { 6000 hammer2_chain_unlock(*parentp); 6001 hammer2_chain_drop(*parentp); 6002 *parentp = NULL; 6003 } 6004 6005 /* 6006 * Be very careful, this is a backend function and we CANNOT 6007 * lock any frontend inode structure we find. But we have to 6008 * look the inode up this way first in case it exists but is 6009 * detached from the radix tree. 6010 */ 6011 ip = hammer2_inode_lookup(pmp, inum); 6012 if (ip) { 6013 *chainp = hammer2_inode_chain_and_parent(ip, clindex, 6014 parentp, 6015 resolve_flags); 6016 hammer2_inode_drop(ip); 6017 if (*chainp) 6018 return (*chainp)->error; 6019 hammer2_chain_unlock(*chainp); 6020 hammer2_chain_drop(*chainp); 6021 *chainp = NULL; 6022 if (*parentp) { 6023 hammer2_chain_unlock(*parentp); 6024 hammer2_chain_drop(*parentp); 6025 *parentp = NULL; 6026 } 6027 } 6028 6029 /* 6030 * Inodes hang off of the iroot (bit 63 is clear, differentiating 6031 * inodes from root directory entries in the key lookup). 6032 */ 6033 parent = hammer2_inode_chain(pmp->iroot, clindex, resolve_flags); 6034 rchain = NULL; 6035 if (parent) { 6036 /* 6037 * NOTE: rchain can be returned as NULL even if error == 0 6038 * (i.e. not found) 6039 */ 6040 rchain = hammer2_chain_lookup(&parent, &key_dummy, 6041 inum, inum, 6042 &error, flags); 6043 /* 6044 * Propagate a chain-specific error to caller. 6045 * 6046 * If the chain is not errored, we must still validate that the inode 6047 * number is correct, because all hell will break loose if it isn't 6048 * correct. It should always be correct so print to the console and 6049 * simulate a CHECK error if it is not. 6050 */ 6051 if (error == 0 && rchain) { 6052 error = rchain->error; 6053 if (error == 0 && rchain->data) { 6054 if (inum != rchain->data->ipdata.meta.inum) { 6055 kprintf("hammer2_chain_inode_find: lookup inum %ld, " 6056 "got valid inode but with inum %ld\n", 6057 inum, rchain->data->ipdata.meta.inum); 6058 error = HAMMER2_ERROR_CHECK; 6059 rchain->error = error; 6060 } 6061 } 6062 } 6063 } else { 6064 error = HAMMER2_ERROR_EIO; 6065 } 6066 *parentp = parent; 6067 *chainp = rchain; 6068 6069 return error; 6070 } 6071 6072 /* 6073 * Used by the bulkscan code to snapshot the synchronized storage for 6074 * a volume, allowing it to be scanned concurrently against normal 6075 * operation. 6076 */ 6077 hammer2_chain_t * 6078 hammer2_chain_bulksnap(hammer2_dev_t *hmp) 6079 { 6080 hammer2_chain_t *copy; 6081 6082 copy = hammer2_chain_alloc(hmp, hmp->spmp, &hmp->vchain.bref); 6083 copy->data = kmalloc(sizeof(copy->data->voldata), 6084 hmp->mmsg, M_WAITOK | M_ZERO); 6085 hammer2_voldata_lock(hmp); 6086 copy->data->voldata = hmp->volsync; 6087 hammer2_voldata_unlock(hmp); 6088 6089 return copy; 6090 } 6091 6092 void 6093 hammer2_chain_bulkdrop(hammer2_chain_t *copy) 6094 { 6095 KKASSERT(copy->bref.type == HAMMER2_BREF_TYPE_VOLUME); 6096 KKASSERT(copy->data); 6097 kfree(copy->data, copy->hmp->mmsg); 6098 copy->data = NULL; 6099 hammer2_chain_drop(copy); 6100 } 6101 6102 /* 6103 * Returns non-zero if the chain (INODE or DIRENT) matches the 6104 * filename. 6105 */ 6106 int 6107 hammer2_chain_dirent_test(hammer2_chain_t *chain, const char *name, 6108 size_t name_len) 6109 { 6110 const hammer2_inode_data_t *ripdata; 6111 6112 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) { 6113 ripdata = &chain->data->ipdata; 6114 if (ripdata->meta.name_len == name_len && 6115 bcmp(ripdata->filename, name, name_len) == 0) { 6116 return 1; 6117 } 6118 } 6119 if (chain->bref.type == HAMMER2_BREF_TYPE_DIRENT && 6120 chain->bref.embed.dirent.namlen == name_len) { 6121 if (name_len > sizeof(chain->bref.check.buf) && 6122 bcmp(chain->data->buf, name, name_len) == 0) { 6123 return 1; 6124 } 6125 if (name_len <= sizeof(chain->bref.check.buf) && 6126 bcmp(chain->bref.check.buf, name, name_len) == 0) { 6127 return 1; 6128 } 6129 } 6130 return 0; 6131 } 6132