1 /* 2 * Copyright (c) 2011-2013 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 * by 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 and hammer2_chain_core structures. 38 * 39 * Chains represent the filesystem media topology in-memory. Any given 40 * chain can represent an inode, indirect block, data, or other types 41 * of blocks. 42 * 43 * This module provides APIs for direct and indirect block searches, 44 * iterations, recursions, creation, deletion, replication, and snapshot 45 * views (used by the flush and snapshot code). 46 * 47 * Generally speaking any modification made to a chain must propagate all 48 * the way back to the volume header, issuing copy-on-write updates to the 49 * blockref tables all the way up. Any chain except the volume header itself 50 * can be flushed to disk at any time, in any order. None of it matters 51 * until we get to the point where we want to synchronize the volume header 52 * (see the flush code). 53 * 54 * The chain structure supports snapshot views in time, which are primarily 55 * used until the related data and meta-data is flushed to allow the 56 * filesystem to make snapshots without requiring it to first flush, 57 * and to allow the filesystem flush and modify the filesystem concurrently 58 * with minimal or no stalls. 59 */ 60 #include <sys/cdefs.h> 61 #include <sys/param.h> 62 #include <sys/systm.h> 63 #include <sys/types.h> 64 #include <sys/lock.h> 65 #include <sys/uuid.h> 66 67 #include "hammer2.h" 68 69 static int hammer2_indirect_optimize; /* XXX SYSCTL */ 70 71 static hammer2_chain_t *hammer2_chain_create_indirect( 72 hammer2_trans_t *trans, hammer2_chain_t *parent, 73 hammer2_key_t key, int keybits, int *errorp); 74 75 /* 76 * We use a red-black tree to guarantee safe lookups under shared locks. 77 * 78 * Chains can be overloaded onto the same index, creating a different 79 * view of a blockref table based on a transaction id. The RBTREE 80 * deconflicts the view by sub-sorting on delete_tid. 81 * 82 * NOTE: Any 'current' chain which is not yet deleted will have a 83 * delete_tid of HAMMER2_MAX_TID (0xFFF....FFFLLU). 84 */ 85 RB_GENERATE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp); 86 87 int 88 hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2) 89 { 90 if (chain1->index < chain2->index) 91 return(-1); 92 if (chain1->index > chain2->index) 93 return(1); 94 if (chain1->delete_tid < chain2->delete_tid) 95 return(-1); 96 if (chain1->delete_tid > chain2->delete_tid) 97 return(1); 98 return(0); 99 } 100 101 /* 102 * Flag chain->parent SUBMODIFIED recursively up to the root. The 103 * recursion can terminate when a parent is encountered with SUBMODIFIED 104 * already set. The flag is NOT set on the passed-in chain. 105 * 106 * This can be confusing because even though chains are multi-homed, 107 * each chain has a specific idea of its parent (chain->parent) which 108 * is singly-homed. 109 * 110 * This flag is used by the flusher's downward recursion to detect 111 * modifications and can only be cleared bottom-up. 112 * 113 * The parent pointer is protected by all the modified children below it 114 * and cannot be changed until they have all been flushed. However, setsubmod 115 * operations on new modifications can race flushes in progress, so we use 116 * the chain->core->cst.spin lock to handle collisions. 117 */ 118 void 119 hammer2_chain_parent_setsubmod(hammer2_chain_t *chain) 120 { 121 hammer2_chain_t *parent; 122 hammer2_chain_core_t *core; 123 124 while ((parent = chain->parent) != NULL) { 125 core = parent->core; 126 spin_lock(&core->cst.spin); 127 if (parent->flags & HAMMER2_CHAIN_SUBMODIFIED) { 128 spin_unlock(&core->cst.spin); 129 break; 130 } 131 atomic_set_int(&parent->flags, HAMMER2_CHAIN_SUBMODIFIED); 132 spin_unlock(&core->cst.spin); 133 chain = parent; 134 } 135 } 136 137 /* 138 * Allocate a new disconnected chain element representing the specified 139 * bref. chain->refs is set to 1 and the passed bref is copied to 140 * chain->bref. chain->bytes is derived from the bref. 141 * 142 * chain->core is NOT allocated and the media data and bp pointers are left 143 * NULL. The caller must call chain_core_alloc() to allocate or associate 144 * a core with the chain. 145 * 146 * NOTE: Returns a referenced but unlocked (because there is no core) chain. 147 */ 148 hammer2_chain_t * 149 hammer2_chain_alloc(hammer2_mount_t *hmp, hammer2_blockref_t *bref) 150 { 151 hammer2_chain_t *chain; 152 u_int bytes = 1U << (int)(bref->data_off & HAMMER2_OFF_MASK_RADIX); 153 154 /* 155 * Construct the appropriate system structure. 156 */ 157 switch(bref->type) { 158 case HAMMER2_BREF_TYPE_INODE: 159 case HAMMER2_BREF_TYPE_INDIRECT: 160 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 161 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 162 case HAMMER2_BREF_TYPE_DATA: 163 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 164 chain = kmalloc(sizeof(*chain), hmp->mchain, M_WAITOK | M_ZERO); 165 break; 166 case HAMMER2_BREF_TYPE_VOLUME: 167 chain = NULL; 168 panic("hammer2_chain_alloc volume type illegal for op"); 169 default: 170 chain = NULL; 171 panic("hammer2_chain_alloc: unrecognized blockref type: %d", 172 bref->type); 173 } 174 175 chain->hmp = hmp; 176 chain->bref = *bref; 177 chain->index = -1; /* not yet assigned */ 178 chain->bytes = bytes; 179 chain->refs = 1; 180 chain->flags = HAMMER2_CHAIN_ALLOCATED; 181 chain->delete_tid = HAMMER2_MAX_TID; 182 183 return (chain); 184 } 185 186 /* 187 * Associate an existing core with the chain or allocate a new core. 188 * 189 * The core is not locked. No additional refs on the chain are made. 190 */ 191 void 192 hammer2_chain_core_alloc(hammer2_chain_t *chain, hammer2_chain_core_t *core) 193 { 194 KKASSERT(chain->core == NULL); 195 196 if (core == NULL) { 197 core = kmalloc(sizeof(*core), chain->hmp->mchain, 198 M_WAITOK | M_ZERO); 199 RB_INIT(&core->rbtree); 200 core->sharecnt = 1; 201 chain->core = core; 202 ccms_cst_init(&core->cst, chain); 203 } else { 204 atomic_add_int(&core->sharecnt, 1); 205 chain->core = core; 206 } 207 } 208 209 /* 210 * Deallocate a chain after the caller has transitioned its refs to 0 211 * and disassociated it from its parent. 212 * 213 * We must drop sharecnt on the core (if any) and handle its 1->0 transition 214 * too. 215 */ 216 static void 217 hammer2_chain_dealloc(hammer2_chain_t *chain) 218 { 219 hammer2_chain_core_t *core; 220 221 /* 222 * Chain's flags are expected to be sane. 223 */ 224 KKASSERT((chain->flags & (HAMMER2_CHAIN_MOVED | 225 HAMMER2_CHAIN_MODIFIED | 226 HAMMER2_CHAIN_ONRBTREE)) == 0); 227 KKASSERT(chain->duplink == NULL); 228 229 /* 230 * Disconnect chain->core from chain and free core if it was the 231 * last core. If any children are present in the core's rbtree 232 * they cannot have a pointer to our chain by definition because 233 * our chain's refs have dropped to 0. If this is the last sharecnt 234 * on core, then core's rbtree must be empty by definition. 235 */ 236 if ((core = chain->core) != NULL) { 237 /* 238 * Other chains may reference the same core so the core's 239 * spinlock is needed to safely disconnect it. 240 */ 241 spin_lock(&core->cst.spin); 242 chain->core = NULL; 243 if (atomic_fetchadd_int(&core->sharecnt, -1) == 1) { 244 spin_unlock(&core->cst.spin); 245 KKASSERT(RB_EMPTY(&core->rbtree)); 246 KKASSERT(core->cst.count == 0); 247 KKASSERT(core->cst.upgrade == 0); 248 kfree(core, chain->hmp->mchain); 249 } else { 250 spin_unlock(&core->cst.spin); 251 } 252 core = NULL; /* safety */ 253 } 254 255 /* 256 * Finally free the structure and return for possible recursion. 257 */ 258 hammer2_chain_free(chain); 259 } 260 261 /* 262 * Free a disconnected chain element. 263 */ 264 void 265 hammer2_chain_free(hammer2_chain_t *chain) 266 { 267 hammer2_mount_t *hmp = chain->hmp; 268 269 switch(chain->bref.type) { 270 case HAMMER2_BREF_TYPE_VOLUME: 271 chain->data = NULL; 272 break; 273 case HAMMER2_BREF_TYPE_INODE: 274 if (chain->data) { 275 kfree(chain->data, hmp->minode); 276 chain->data = NULL; 277 } 278 break; 279 default: 280 KKASSERT(chain->data == NULL); 281 break; 282 } 283 284 KKASSERT(chain->core == NULL); 285 KKASSERT(chain->bp == NULL); 286 chain->hmp = NULL; 287 288 if (chain->flags & HAMMER2_CHAIN_ALLOCATED) 289 kfree(chain, hmp->mchain); 290 } 291 292 /* 293 * Add a reference to a chain element, preventing its destruction. 294 */ 295 void 296 hammer2_chain_ref(hammer2_chain_t *chain) 297 { 298 atomic_add_int(&chain->refs, 1); 299 } 300 301 /* 302 * Drop the caller's reference to the chain. When the ref count drops to 303 * zero this function will disassociate the chain from its parent and 304 * deallocate it, then recursely drop the parent using the implied ref 305 * from the chain's chain->parent. 306 * 307 * WARNING! Just because we are able to deallocate a chain doesn't mean 308 * that chain->core->rbtree is empty. There can still be a sharecnt 309 * on chain->core and RBTREE entries that refer to different parents. 310 */ 311 static hammer2_chain_t *hammer2_chain_lastdrop(hammer2_chain_t *chain); 312 313 void 314 hammer2_chain_drop(hammer2_chain_t *chain) 315 { 316 u_int refs; 317 318 while (chain) { 319 refs = chain->refs; 320 cpu_ccfence(); 321 KKASSERT(refs > 0); 322 323 if (refs == 1) { 324 if (chain->parent) { 325 chain = hammer2_chain_lastdrop(chain); 326 /* recursively drop parent or retry same */ 327 } else if (atomic_cmpset_int(&chain->refs, 1, 0)) { 328 hammer2_chain_dealloc(chain); 329 chain = NULL; 330 /* no parent to recurse on */ 331 } else { 332 /* retry the same chain */ 333 } 334 } else { 335 if (atomic_cmpset_int(&chain->refs, refs, refs - 1)) 336 break; 337 /* retry the same chain */ 338 } 339 } 340 } 341 342 /* 343 * Safe handling of the 1->0 transition on chain when the chain has a 344 * parent. 345 * 346 * NOTE: A chain can only be removed from its parent core's RBTREE on 347 * the 1->0 transition by definition. No other code is allowed 348 * to remove chain from its RBTREE, so no race is possible. 349 */ 350 static 351 hammer2_chain_t * 352 hammer2_chain_lastdrop(hammer2_chain_t *chain) 353 { 354 hammer2_chain_t *parent; 355 hammer2_chain_core_t *parent_core; 356 357 parent = chain->parent; 358 parent_core = parent->core; 359 KKASSERT(chain->flags & HAMMER2_CHAIN_ONRBTREE); 360 361 spin_lock(&parent_core->cst.spin); 362 if (atomic_cmpset_int(&chain->refs, 1, 0)) { 363 RB_REMOVE(hammer2_chain_tree, &parent_core->rbtree, chain); 364 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 365 chain->parent = NULL; /* NULL field, must drop implied ref */ 366 spin_unlock(&parent_core->cst.spin); 367 if (chain->duplink) { 368 hammer2_chain_drop(chain->duplink); 369 chain->duplink = NULL; 370 } 371 hammer2_chain_dealloc(chain); 372 chain = parent; /* recursively drop parent */ 373 } else { 374 spin_unlock(&parent_core->cst.spin); 375 } 376 return (chain); 377 } 378 379 /* 380 * Ref and lock a chain element, acquiring its data with I/O if necessary, 381 * and specify how you would like the data to be resolved. 382 * 383 * Returns 0 on success or an error code if the data could not be acquired. 384 * The chain element is locked either way. 385 * 386 * The lock is allowed to recurse, multiple locking ops will aggregate 387 * the requested resolve types. Once data is assigned it will not be 388 * removed until the last unlock. 389 * 390 * HAMMER2_RESOLVE_NEVER - Do not resolve the data element. 391 * (typically used to avoid device/logical buffer 392 * aliasing for data) 393 * 394 * HAMMER2_RESOLVE_MAYBE - Do not resolve data elements for chains in 395 * the INITIAL-create state (indirect blocks only). 396 * 397 * Do not resolve data elements for DATA chains. 398 * (typically used to avoid device/logical buffer 399 * aliasing for data) 400 * 401 * HAMMER2_RESOLVE_ALWAYS- Always resolve the data element. 402 * 403 * HAMMER2_RESOLVE_SHARED- (flag) The chain is locked shared, otherwise 404 * it will be locked exclusive. 405 * 406 * NOTE: Embedded elements (volume header, inodes) are always resolved 407 * regardless. 408 * 409 * NOTE: Specifying HAMMER2_RESOLVE_ALWAYS on a newly-created non-embedded 410 * element will instantiate and zero its buffer, and flush it on 411 * release. 412 * 413 * NOTE: (data) elements are normally locked RESOLVE_NEVER or RESOLVE_MAYBE 414 * so as not to instantiate a device buffer, which could alias against 415 * a logical file buffer. However, if ALWAYS is specified the 416 * device buffer will be instantiated anyway. 417 * 418 * WARNING! If data must be fetched a shared lock will temporarily be 419 * upgraded to exclusive. However, a deadlock can occur if 420 * the caller owns more than one shared lock. 421 */ 422 int 423 hammer2_chain_lock(hammer2_chain_t *chain, int how) 424 { 425 hammer2_mount_t *hmp; 426 hammer2_chain_core_t *core; 427 hammer2_blockref_t *bref; 428 hammer2_off_t pbase; 429 hammer2_off_t peof; 430 ccms_state_t ostate; 431 size_t boff; 432 size_t bbytes; 433 int error; 434 char *bdata; 435 436 /* 437 * Ref and lock the element. Recursive locks are allowed. 438 */ 439 if ((how & HAMMER2_RESOLVE_NOREF) == 0) 440 hammer2_chain_ref(chain); 441 hmp = chain->hmp; 442 KKASSERT(hmp != NULL); 443 444 /* 445 * Get the appropriate lock. 446 */ 447 core = chain->core; 448 if (how & HAMMER2_RESOLVE_SHARED) 449 ccms_thread_lock(&core->cst, CCMS_STATE_SHARED); 450 else 451 ccms_thread_lock(&core->cst, CCMS_STATE_EXCLUSIVE); 452 453 /* 454 * If we already have a valid data pointer no further action is 455 * necessary. 456 */ 457 if (chain->data) 458 return (0); 459 460 /* 461 * Do we have to resolve the data? 462 */ 463 switch(how & HAMMER2_RESOLVE_MASK) { 464 case HAMMER2_RESOLVE_NEVER: 465 return(0); 466 case HAMMER2_RESOLVE_MAYBE: 467 if (chain->flags & HAMMER2_CHAIN_INITIAL) 468 return(0); 469 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) 470 return(0); 471 if (chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_LEAF) 472 return(0); 473 /* fall through */ 474 case HAMMER2_RESOLVE_ALWAYS: 475 break; 476 } 477 478 /* 479 * Upgrade to an exclusive lock so we can safely manipulate the 480 * buffer cache. If another thread got to it before us we 481 * can just return. 482 */ 483 ostate = ccms_thread_lock_upgrade(&core->cst); 484 if (chain->data) { 485 ccms_thread_lock_restore(&core->cst, ostate); 486 return (0); 487 } 488 489 /* 490 * We must resolve to a device buffer, either by issuing I/O or 491 * by creating a zero-fill element. We do not mark the buffer 492 * dirty when creating a zero-fill element (the hammer2_chain_modify() 493 * API must still be used to do that). 494 * 495 * The device buffer is variable-sized in powers of 2 down 496 * to HAMMER2_MINALLOCSIZE (typically 1K). A 64K physical storage 497 * chunk always contains buffers of the same size. (XXX) 498 * 499 * The minimum physical IO size may be larger than the variable 500 * block size. 501 */ 502 bref = &chain->bref; 503 504 if ((bbytes = chain->bytes) < HAMMER2_MINIOSIZE) 505 bbytes = HAMMER2_MINIOSIZE; 506 pbase = bref->data_off & ~(hammer2_off_t)(bbytes - 1); 507 peof = (pbase + HAMMER2_PBUFSIZE64) & ~HAMMER2_PBUFMASK64; 508 boff = bref->data_off & HAMMER2_OFF_MASK & (bbytes - 1); 509 KKASSERT(pbase != 0); 510 511 /* 512 * The getblk() optimization can only be used on newly created 513 * elements if the physical block size matches the request. 514 */ 515 if ((chain->flags & HAMMER2_CHAIN_INITIAL) && 516 chain->bytes == bbytes) { 517 chain->bp = getblk(hmp->devvp, pbase, bbytes, 0, 0); 518 error = 0; 519 } else if (hammer2_cluster_enable) { 520 error = cluster_read(hmp->devvp, peof, pbase, bbytes, 521 HAMMER2_PBUFSIZE, HAMMER2_PBUFSIZE, 522 &chain->bp); 523 } else { 524 error = bread(hmp->devvp, pbase, bbytes, &chain->bp); 525 } 526 527 if (error) { 528 kprintf("hammer2_chain_get: I/O error %016jx: %d\n", 529 (intmax_t)pbase, error); 530 bqrelse(chain->bp); 531 chain->bp = NULL; 532 ccms_thread_lock_restore(&core->cst, ostate); 533 return (error); 534 } 535 536 /* 537 * Zero the data area if the chain is in the INITIAL-create state. 538 * Mark the buffer for bdwrite(). 539 */ 540 bdata = (char *)chain->bp->b_data + boff; 541 if (chain->flags & HAMMER2_CHAIN_INITIAL) { 542 bzero(bdata, chain->bytes); 543 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DIRTYBP); 544 } 545 546 /* 547 * Setup the data pointer, either pointing it to an embedded data 548 * structure and copying the data from the buffer, or pointing it 549 * into the buffer. 550 * 551 * The buffer is not retained when copying to an embedded data 552 * structure in order to avoid potential deadlocks or recursions 553 * on the same physical buffer. 554 */ 555 switch (bref->type) { 556 case HAMMER2_BREF_TYPE_VOLUME: 557 /* 558 * Copy data from bp to embedded buffer 559 */ 560 panic("hammer2_chain_lock: called on unresolved volume header"); 561 #if 0 562 /* NOT YET */ 563 KKASSERT(pbase == 0); 564 KKASSERT(chain->bytes == HAMMER2_PBUFSIZE); 565 bcopy(bdata, &hmp->voldata, chain->bytes); 566 chain->data = (void *)&hmp->voldata; 567 bqrelse(chain->bp); 568 chain->bp = NULL; 569 #endif 570 break; 571 case HAMMER2_BREF_TYPE_INODE: 572 /* 573 * Copy data from bp to embedded buffer, do not retain the 574 * device buffer. 575 */ 576 KKASSERT(chain->bytes == sizeof(chain->data->ipdata)); 577 chain->data = kmalloc(sizeof(chain->data->ipdata), 578 hmp->minode, M_WAITOK | M_ZERO); 579 bcopy(bdata, &chain->data->ipdata, chain->bytes); 580 bqrelse(chain->bp); 581 chain->bp = NULL; 582 break; 583 case HAMMER2_BREF_TYPE_INDIRECT: 584 case HAMMER2_BREF_TYPE_DATA: 585 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 586 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 587 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 588 default: 589 /* 590 * Point data at the device buffer and leave bp intact. 591 */ 592 chain->data = (void *)bdata; 593 break; 594 } 595 596 /* 597 * Make sure the bp is not specifically owned by this thread before 598 * restoring to a possibly shared lock, so another hammer2 thread 599 * can release it. 600 */ 601 if (chain->bp) 602 BUF_KERNPROC(chain->bp); 603 ccms_thread_lock_restore(&core->cst, ostate); 604 return (0); 605 } 606 607 /* 608 * Unlock and deref a chain element. 609 * 610 * On the last lock release any non-embedded data (chain->bp) will be 611 * retired. 612 */ 613 void 614 hammer2_chain_unlock(hammer2_chain_t *chain) 615 { 616 hammer2_chain_core_t *core = chain->core; 617 long *counterp; 618 619 /* 620 * Release the CST lock but with a special 1->0 transition case 621 * to also drop the refs on chain. Multiple CST locks only 622 * 623 * Returns non-zero if lock references remain. When zero is 624 * returned the last lock reference is retained and any shared 625 * lock is upgraded to an exclusive lock for final disposition. 626 */ 627 if (ccms_thread_unlock_zero(&core->cst)) { 628 KKASSERT(chain->refs > 1); 629 atomic_add_int(&chain->refs, -1); 630 return; 631 } 632 633 /* 634 * Shortcut the case if the data is embedded or not resolved. 635 * 636 * Do NOT NULL out chain->data (e.g. inode data), it might be 637 * dirty. 638 * 639 * The DIRTYBP flag is non-applicable in this situation and can 640 * be cleared to keep the flags state clean. 641 */ 642 if (chain->bp == NULL) { 643 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DIRTYBP); 644 ccms_thread_unlock(&core->cst); 645 hammer2_chain_drop(chain); 646 return; 647 } 648 649 /* 650 * Statistics 651 */ 652 if ((chain->flags & HAMMER2_CHAIN_DIRTYBP) == 0) { 653 ; 654 } else if (chain->flags & HAMMER2_CHAIN_IOFLUSH) { 655 switch(chain->bref.type) { 656 case HAMMER2_BREF_TYPE_DATA: 657 counterp = &hammer2_ioa_file_write; 658 break; 659 case HAMMER2_BREF_TYPE_INODE: 660 counterp = &hammer2_ioa_meta_write; 661 break; 662 case HAMMER2_BREF_TYPE_INDIRECT: 663 counterp = &hammer2_ioa_indr_write; 664 break; 665 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 666 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 667 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 668 counterp = &hammer2_ioa_fmap_write; 669 break; 670 default: 671 counterp = &hammer2_ioa_volu_write; 672 break; 673 } 674 ++*counterp; 675 } else { 676 switch(chain->bref.type) { 677 case HAMMER2_BREF_TYPE_DATA: 678 counterp = &hammer2_iod_file_write; 679 break; 680 case HAMMER2_BREF_TYPE_INODE: 681 counterp = &hammer2_iod_meta_write; 682 break; 683 case HAMMER2_BREF_TYPE_INDIRECT: 684 counterp = &hammer2_iod_indr_write; 685 break; 686 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 687 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 688 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 689 counterp = &hammer2_iod_fmap_write; 690 break; 691 default: 692 counterp = &hammer2_iod_volu_write; 693 break; 694 } 695 ++*counterp; 696 } 697 698 /* 699 * Clean out the bp. 700 * 701 * If a device buffer was used for data be sure to destroy the 702 * buffer when we are done to avoid aliases (XXX what about the 703 * underlying VM pages?). 704 * 705 * NOTE: Freemap leaf's use reserved blocks and thus no aliasing 706 * is possible. 707 */ 708 if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) 709 chain->bp->b_flags |= B_RELBUF; 710 711 /* 712 * The DIRTYBP flag tracks whether we have to bdwrite() the buffer 713 * or not. The flag will get re-set when chain_modify() is called, 714 * even if MODIFIED is already set, allowing the OS to retire the 715 * buffer independent of a hammer2 flus. 716 */ 717 chain->data = NULL; 718 if (chain->flags & HAMMER2_CHAIN_DIRTYBP) { 719 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_DIRTYBP); 720 if (chain->flags & HAMMER2_CHAIN_IOFLUSH) { 721 atomic_clear_int(&chain->flags, 722 HAMMER2_CHAIN_IOFLUSH); 723 chain->bp->b_flags |= B_RELBUF; 724 cluster_awrite(chain->bp); 725 } else { 726 chain->bp->b_flags |= B_CLUSTEROK; 727 bdwrite(chain->bp); 728 } 729 } else { 730 if (chain->flags & HAMMER2_CHAIN_IOFLUSH) { 731 atomic_clear_int(&chain->flags, 732 HAMMER2_CHAIN_IOFLUSH); 733 chain->bp->b_flags |= B_RELBUF; 734 brelse(chain->bp); 735 } else { 736 /* bp might still be dirty */ 737 bqrelse(chain->bp); 738 } 739 } 740 chain->bp = NULL; 741 ccms_thread_unlock(&core->cst); 742 hammer2_chain_drop(chain); 743 } 744 745 /* 746 * Resize the chain's physical storage allocation in-place. This may 747 * replace the passed-in chain with a new chain. 748 * 749 * Chains can be resized smaller without reallocating the storage. 750 * Resizing larger will reallocate the storage. 751 * 752 * Must be passed an exclusively locked parent and chain, returns a new 753 * exclusively locked chain at the same index and unlocks the old chain. 754 * Flushes the buffer if necessary. 755 * 756 * If you want the resize code to copy the data to the new block then the 757 * caller should lock the chain RESOLVE_MAYBE or RESOLVE_ALWAYS. 758 * 759 * If the caller already holds a logical buffer containing the data and 760 * intends to bdwrite() that buffer resolve with RESOLVE_NEVER. The resize 761 * operation will then not copy the (stale) data from the media. 762 * 763 * This function is mostly used with DATA blocks locked RESOLVE_NEVER in order 764 * to avoid instantiating a device buffer that conflicts with the vnode 765 * data buffer. 766 * 767 * XXX flags currently ignored, uses chain->bp to detect data/no-data. 768 * XXX return error if cannot resize. 769 */ 770 void 771 hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip, 772 struct buf *bp, 773 hammer2_chain_t *parent, hammer2_chain_t **chainp, 774 int nradix, int flags) 775 { 776 hammer2_mount_t *hmp = trans->hmp; 777 hammer2_chain_t *chain = *chainp; 778 struct buf *nbp; 779 hammer2_off_t pbase; 780 size_t obytes; 781 size_t nbytes; 782 size_t bbytes; 783 int boff; 784 char *bdata; 785 int error; 786 787 /* 788 * Only data and indirect blocks can be resized for now. 789 * (The volu root, inodes, and freemap elements use a fixed size). 790 */ 791 KKASSERT(chain != &hmp->vchain); 792 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_DATA || 793 chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT); 794 795 /* 796 * Nothing to do if the element is already the proper size 797 */ 798 obytes = chain->bytes; 799 nbytes = 1U << nradix; 800 if (obytes == nbytes) 801 return; 802 803 /* 804 * Delete the old chain and duplicate it at the same (parent, index), 805 * returning a new chain. This allows the old chain to still be 806 * used by the flush code. Duplication occurs in-place. 807 * 808 * NOTE: If we are not crossing a synchronization point the 809 * duplication code will simply reuse the existing chain 810 * structure. 811 */ 812 hammer2_chain_delete(trans, parent, chain); 813 hammer2_chain_duplicate(trans, parent, chain->index, &chain); 814 815 /* 816 * Set MODIFIED and add a chain ref to prevent destruction. Both 817 * modified flags share the same ref. (duplicated chains do not 818 * start out MODIFIED unless possibly if the duplication code 819 * decided to reuse the existing chain as-is). 820 * 821 * If the chain is already marked MODIFIED then we can safely 822 * return the previous allocation to the pool without having to 823 * worry about snapshots. 824 */ 825 if ((chain->flags & HAMMER2_CHAIN_MODIFIED) == 0) { 826 atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED); 827 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 828 hammer2_chain_ref(chain); 829 } else { 830 hammer2_freemap_free(hmp, chain->bref.data_off, 831 chain->bref.type); 832 } 833 834 /* 835 * Relocate the block, even if making it smaller (because different 836 * block sizes may be in different regions). 837 */ 838 chain->bref.data_off = hammer2_freemap_alloc(hmp, chain->bref.type, 839 nbytes); 840 chain->bytes = nbytes; 841 /*ip->delta_dcount += (ssize_t)(nbytes - obytes);*/ /* XXX atomic */ 842 843 /* 844 * The device buffer may be larger than the allocation size. 845 */ 846 if ((bbytes = chain->bytes) < HAMMER2_MINIOSIZE) 847 bbytes = HAMMER2_MINIOSIZE; 848 pbase = chain->bref.data_off & ~(hammer2_off_t)(bbytes - 1); 849 boff = chain->bref.data_off & HAMMER2_OFF_MASK & (bbytes - 1); 850 851 /* 852 * Only copy the data if resolved, otherwise the caller is 853 * responsible. 854 */ 855 if (chain->bp) { 856 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 857 chain->bref.type == HAMMER2_BREF_TYPE_DATA); 858 KKASSERT(chain != &hmp->vchain); /* safety */ 859 860 /* 861 * The getblk() optimization can only be used if the 862 * physical block size matches the request. 863 */ 864 if (nbytes == bbytes) { 865 nbp = getblk(hmp->devvp, pbase, bbytes, 0, 0); 866 error = 0; 867 } else { 868 error = bread(hmp->devvp, pbase, bbytes, &nbp); 869 KKASSERT(error == 0); 870 } 871 bdata = (char *)nbp->b_data + boff; 872 873 /* 874 * chain->bp and chain->data represent the on-disk version 875 * of the data, where as the passed-in bp is usually a 876 * more up-to-date logical buffer. However, there is no 877 * need to synchronize the more up-to-date data in (bp) 878 * as it will do that on its own when it flushes. 879 */ 880 if (nbytes < obytes) { 881 bcopy(chain->data, bdata, nbytes); 882 } else { 883 bcopy(chain->data, bdata, obytes); 884 bzero(bdata + obytes, nbytes - obytes); 885 } 886 887 /* 888 * NOTE: The INITIAL state of the chain is left intact. 889 * We depend on hammer2_chain_modify() to do the 890 * right thing. 891 * 892 * NOTE: We set B_NOCACHE to throw away the previous bp and 893 * any VM backing store, even if it was dirty. 894 * Otherwise we run the risk of a logical/device 895 * conflict on reallocation. 896 */ 897 chain->bp->b_flags |= B_RELBUF | B_NOCACHE; 898 brelse(chain->bp); 899 chain->bp = nbp; 900 chain->data = (void *)bdata; 901 hammer2_chain_modify(trans, chain, 0); 902 } 903 904 /* 905 * Make sure the chain is marked MOVED and SUBMOD is set in the 906 * parent(s) so the adjustments are picked up by flush. 907 */ 908 if ((chain->flags & HAMMER2_CHAIN_MOVED) == 0) { 909 hammer2_chain_ref(chain); 910 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MOVED); 911 } 912 hammer2_chain_parent_setsubmod(chain); 913 } 914 915 /* 916 * Convert a locked chain that was retrieved read-only to read-write. 917 * 918 * If not already marked modified a new physical block will be allocated 919 * and assigned to the bref. 920 * 921 * If already modified and the new modification crosses a synchronization 922 * point the chain is duplicated in order to allow the flush to synchronize 923 * the old chain. The new chain replaces the old. 924 * 925 * Non-data blocks - The chain should be locked to at least the RESOLVE_MAYBE 926 * level or the COW operation will not work. 927 * 928 * Data blocks - The chain is usually locked RESOLVE_NEVER so as not to 929 * run the data through the device buffers. 930 * 931 * This function may return a different chain than was passed, in which case 932 * the old chain will be unlocked and the new chain will be locked. 933 */ 934 void 935 hammer2_chain_modify(hammer2_trans_t *trans, hammer2_chain_t *chain, int flags) 936 { 937 hammer2_mount_t *hmp = trans->hmp; 938 hammer2_off_t pbase; 939 struct buf *nbp; 940 int error; 941 size_t bbytes; 942 size_t boff; 943 void *bdata; 944 945 /* 946 * modify_tid is only update for primary modifications, not for 947 * propagated brefs. mirror_tid will be updated regardless during 948 * the flush, no need to set it here. 949 */ 950 if ((flags & HAMMER2_MODIFY_NO_MODIFY_TID) == 0) 951 chain->bref.modify_tid = trans->sync_tid; 952 953 /* 954 * If the chain is already marked MODIFIED we can just return. 955 * 956 * However, it is possible that a prior lock/modify sequence 957 * retired the buffer. During this lock/modify sequence MODIFIED 958 * may still be set but the buffer could wind up clean. Since 959 * the caller is going to modify the buffer further we have to 960 * be sure that DIRTYBP is set again. 961 */ 962 if (chain->flags & HAMMER2_CHAIN_MODIFIED) { 963 if ((flags & HAMMER2_MODIFY_OPTDATA) == 0 && 964 chain->bp == NULL) { 965 goto skip1; 966 } 967 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DIRTYBP); 968 return; 969 } 970 971 /* 972 * Set MODIFIED and add a chain ref to prevent destruction. Both 973 * modified flags share the same ref. 974 */ 975 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MODIFIED); 976 hammer2_chain_ref(chain); 977 978 /* 979 * We must allocate the copy-on-write block. 980 * 981 * If the data is embedded no other action is required. 982 * 983 * If the data is not embedded we acquire and clear the 984 * new block. If chain->data is not NULL we then do the 985 * copy-on-write. chain->data will then be repointed to the new 986 * buffer and the old buffer will be released. 987 * 988 * For newly created elements with no prior allocation we go 989 * through the copy-on-write steps except without the copying part. 990 */ 991 if (chain != &hmp->vchain) { 992 if ((hammer2_debug & 0x0001) && 993 (chain->bref.data_off & HAMMER2_OFF_MASK)) { 994 kprintf("Replace %d\n", chain->bytes); 995 } 996 chain->bref.data_off = 997 hammer2_freemap_alloc(hmp, chain->bref.type, 998 chain->bytes); 999 /* XXX failed allocation */ 1000 } 1001 1002 /* 1003 * If data instantiation is optional and the chain has no current 1004 * data association (typical for DATA and newly-created INDIRECT 1005 * elements), don't instantiate the buffer now. 1006 */ 1007 if ((flags & HAMMER2_MODIFY_OPTDATA) && chain->bp == NULL) 1008 goto skip2; 1009 1010 skip1: 1011 /* 1012 * Setting the DIRTYBP flag will cause the buffer to be dirtied or 1013 * written-out on unlock. This bit is independent of the MODIFIED 1014 * bit because the chain may still need meta-data adjustments done 1015 * by virtue of MODIFIED for its parent, and the buffer can be 1016 * flushed out (possibly multiple times) by the OS before that. 1017 * 1018 * Clearing the INITIAL flag (for indirect blocks) indicates that 1019 * a zero-fill buffer has been instantiated. 1020 */ 1021 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DIRTYBP); 1022 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1023 1024 /* 1025 * We currently should never instantiate a device buffer for a 1026 * file data chain. (We definitely can for a freemap chain). 1027 */ 1028 KKASSERT(chain->bref.type != HAMMER2_BREF_TYPE_DATA); 1029 1030 /* 1031 * Execute COW operation 1032 */ 1033 switch(chain->bref.type) { 1034 case HAMMER2_BREF_TYPE_VOLUME: 1035 case HAMMER2_BREF_TYPE_INODE: 1036 /* 1037 * The data is embedded, no copy-on-write operation is 1038 * needed. 1039 */ 1040 KKASSERT(chain->bp == NULL); 1041 break; 1042 case HAMMER2_BREF_TYPE_DATA: 1043 case HAMMER2_BREF_TYPE_INDIRECT: 1044 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1045 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1046 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1047 /* 1048 * Perform the copy-on-write operation 1049 */ 1050 KKASSERT(chain != &hmp->vchain); /* safety */ 1051 /* 1052 * The device buffer may be larger than the allocation size. 1053 */ 1054 if ((bbytes = chain->bytes) < HAMMER2_MINIOSIZE) 1055 bbytes = HAMMER2_MINIOSIZE; 1056 pbase = chain->bref.data_off & ~(hammer2_off_t)(bbytes - 1); 1057 boff = chain->bref.data_off & HAMMER2_OFF_MASK & (bbytes - 1); 1058 1059 /* 1060 * The getblk() optimization can only be used if the 1061 * physical block size matches the request. 1062 */ 1063 if (chain->bytes == bbytes) { 1064 nbp = getblk(hmp->devvp, pbase, bbytes, 0, 0); 1065 error = 0; 1066 } else { 1067 error = bread(hmp->devvp, pbase, bbytes, &nbp); 1068 KKASSERT(error == 0); 1069 } 1070 bdata = (char *)nbp->b_data + boff; 1071 1072 /* 1073 * Copy or zero-fill on write depending on whether 1074 * chain->data exists or not. 1075 */ 1076 if (chain->data) { 1077 bcopy(chain->data, bdata, chain->bytes); 1078 KKASSERT(chain->bp != NULL); 1079 } else { 1080 bzero(bdata, chain->bytes); 1081 } 1082 if (chain->bp) { 1083 chain->bp->b_flags |= B_RELBUF; 1084 brelse(chain->bp); 1085 } 1086 chain->bp = nbp; 1087 chain->data = bdata; 1088 break; 1089 default: 1090 panic("hammer2_chain_modify: illegal non-embedded type %d", 1091 chain->bref.type); 1092 break; 1093 1094 } 1095 skip2: 1096 if ((flags & HAMMER2_MODIFY_NOSUB) == 0) 1097 hammer2_chain_parent_setsubmod(chain); 1098 } 1099 1100 /* 1101 * Mark the volume as having been modified. This short-cut version 1102 * does not have to lock the volume's chain, which allows the ioctl 1103 * code to make adjustments to connections without deadlocking. XXX 1104 * 1105 * No ref is made on vchain when flagging it MODIFIED. 1106 */ 1107 void 1108 hammer2_modify_volume(hammer2_mount_t *hmp) 1109 { 1110 hammer2_voldata_lock(hmp); 1111 hammer2_voldata_unlock(hmp, 1); 1112 } 1113 1114 /* 1115 * Locate an in-memory chain. The parent must be locked. The in-memory 1116 * chain is returned with a reference and without a lock, or NULL 1117 * if not found. 1118 * 1119 * NOTE: A chain on-media might exist for this index when NULL is returned. 1120 * 1121 * NOTE: Can only be used to locate chains which have not been deleted. 1122 */ 1123 hammer2_chain_t * 1124 hammer2_chain_find(hammer2_chain_t *parent, int index) 1125 { 1126 hammer2_chain_t dummy; 1127 hammer2_chain_t *chain; 1128 1129 dummy.flags = 0; 1130 dummy.index = index; 1131 dummy.delete_tid = HAMMER2_MAX_TID; 1132 spin_lock(&parent->core->cst.spin); 1133 chain = RB_FIND(hammer2_chain_tree, &parent->core->rbtree, &dummy); 1134 if (chain) 1135 hammer2_chain_ref(chain); 1136 spin_unlock(&parent->core->cst.spin); 1137 1138 return (chain); 1139 } 1140 1141 /* 1142 * Return a locked chain structure with all associated data acquired. 1143 * (if LOOKUP_NOLOCK is requested the returned chain is only referenced). 1144 * 1145 * Caller must hold the parent locked shared or exclusive since we may 1146 * need the parent's bref array to find our block. 1147 * 1148 * The returned child is locked as requested. If NOLOCK, the returned 1149 * child is still at least referenced. 1150 */ 1151 hammer2_chain_t * 1152 hammer2_chain_get(hammer2_chain_t *parent, int index, int flags) 1153 { 1154 hammer2_blockref_t *bref; 1155 hammer2_mount_t *hmp = parent->hmp; 1156 hammer2_chain_t *chain; 1157 hammer2_chain_t dummy; 1158 int how; 1159 1160 /* 1161 * Figure out how to lock. MAYBE can be used to optimized 1162 * the initial-create state for indirect blocks. 1163 */ 1164 if (flags & (HAMMER2_LOOKUP_NODATA | HAMMER2_LOOKUP_NOLOCK)) 1165 how = HAMMER2_RESOLVE_NEVER; 1166 else 1167 how = HAMMER2_RESOLVE_MAYBE; 1168 if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK)) 1169 how |= HAMMER2_RESOLVE_SHARED; 1170 1171 retry: 1172 /* 1173 * First see if we have a (possibly modified) chain element cached 1174 * for this (parent, index). Acquire the data if necessary. 1175 * 1176 * If chain->data is non-NULL the chain should already be marked 1177 * modified. 1178 */ 1179 dummy.flags = 0; 1180 dummy.index = index; 1181 dummy.delete_tid = HAMMER2_MAX_TID; 1182 spin_lock(&parent->core->cst.spin); 1183 chain = RB_FIND(hammer2_chain_tree, &parent->core->rbtree, &dummy); 1184 if (chain) { 1185 hammer2_chain_ref(chain); 1186 spin_unlock(&parent->core->cst.spin); 1187 if ((flags & HAMMER2_LOOKUP_NOLOCK) == 0) 1188 hammer2_chain_lock(chain, how | HAMMER2_RESOLVE_NOREF); 1189 return(chain); 1190 } 1191 spin_unlock(&parent->core->cst.spin); 1192 1193 /* 1194 * The parent chain must not be in the INITIAL state. 1195 */ 1196 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 1197 panic("hammer2_chain_get: Missing bref(1)"); 1198 /* NOT REACHED */ 1199 } 1200 1201 /* 1202 * No RBTREE entry found, lookup the bref and issue I/O (switch on 1203 * the parent's bref to determine where and how big the array is). 1204 */ 1205 switch(parent->bref.type) { 1206 case HAMMER2_BREF_TYPE_INODE: 1207 KKASSERT(index >= 0 && index < HAMMER2_SET_COUNT); 1208 bref = &parent->data->ipdata.u.blockset.blockref[index]; 1209 break; 1210 case HAMMER2_BREF_TYPE_INDIRECT: 1211 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1212 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1213 KKASSERT(parent->data != NULL); 1214 KKASSERT(index >= 0 && 1215 index < parent->bytes / sizeof(hammer2_blockref_t)); 1216 bref = &parent->data->npdata.blockref[index]; 1217 break; 1218 case HAMMER2_BREF_TYPE_VOLUME: 1219 KKASSERT(index >= 0 && index < HAMMER2_SET_COUNT); 1220 bref = &hmp->voldata.sroot_blockset.blockref[index]; 1221 break; 1222 default: 1223 bref = NULL; 1224 panic("hammer2_chain_get: unrecognized blockref type: %d", 1225 parent->bref.type); 1226 } 1227 if (bref->type == 0) { 1228 panic("hammer2_chain_get: Missing bref(2)"); 1229 /* NOT REACHED */ 1230 } 1231 1232 /* 1233 * Allocate a chain structure representing the existing media 1234 * entry. Resulting chain has one ref and is not locked. 1235 * 1236 * The locking operation we do later will issue I/O to read it. 1237 */ 1238 chain = hammer2_chain_alloc(hmp, bref); 1239 hammer2_chain_core_alloc(chain, NULL); /* ref'd chain returned */ 1240 1241 /* 1242 * Link the chain into its parent. A spinlock is required to safely 1243 * access the RBTREE, and it is possible to collide with another 1244 * hammer2_chain_get() operation because the caller might only hold 1245 * a shared lock on the parent. 1246 */ 1247 KKASSERT(parent->refs > 0); 1248 spin_lock(&parent->core->cst.spin); 1249 chain->parent = parent; 1250 chain->index = index; 1251 if (RB_INSERT(hammer2_chain_tree, &parent->core->rbtree, chain)) { 1252 chain->parent = NULL; 1253 chain->index = -1; 1254 spin_unlock(&parent->core->cst.spin); 1255 hammer2_chain_drop(chain); 1256 goto retry; 1257 } 1258 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 1259 hammer2_chain_ref(parent); /* chain->parent ref */ 1260 spin_unlock(&parent->core->cst.spin); 1261 1262 /* 1263 * Our new chain is referenced but NOT locked. Lock the chain 1264 * below. The locking operation also resolves its data. 1265 * 1266 * If NOLOCK is set the release will release the one-and-only lock. 1267 */ 1268 if ((flags & HAMMER2_LOOKUP_NOLOCK) == 0) { 1269 hammer2_chain_lock(chain, how); /* recusive lock */ 1270 hammer2_chain_drop(chain); /* excess ref */ 1271 } 1272 return (chain); 1273 } 1274 1275 /* 1276 * Lookup initialization/completion API 1277 */ 1278 hammer2_chain_t * 1279 hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags) 1280 { 1281 if (flags & HAMMER2_LOOKUP_SHARED) { 1282 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS | 1283 HAMMER2_RESOLVE_SHARED); 1284 } else { 1285 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS); 1286 } 1287 return (parent); 1288 } 1289 1290 void 1291 hammer2_chain_lookup_done(hammer2_chain_t *parent) 1292 { 1293 if (parent) 1294 hammer2_chain_unlock(parent); 1295 } 1296 1297 1298 /* 1299 * Locate any key between key_beg and key_end inclusive. (*parentp) 1300 * typically points to an inode but can also point to a related indirect 1301 * block and this function will recurse upwards and find the inode again. 1302 * 1303 * WARNING! THIS DOES NOT RETURN KEYS IN LOGICAL KEY ORDER! ANY KEY 1304 * WITHIN THE RANGE CAN BE RETURNED. HOWEVER, AN ITERATION 1305 * WHICH PICKS UP WHERE WE LEFT OFF WILL CONTINUE THE SCAN. 1306 * 1307 * (*parentp) must be exclusively locked and referenced and can be an inode 1308 * or an existing indirect block within the inode. 1309 * 1310 * On return (*parentp) will be modified to point at the deepest parent chain 1311 * element encountered during the search, as a helper for an insertion or 1312 * deletion. The new (*parentp) will be locked and referenced and the old 1313 * will be unlocked and dereferenced (no change if they are both the same). 1314 * 1315 * The matching chain will be returned exclusively locked. If NOLOCK is 1316 * requested the chain will be returned only referenced. 1317 * 1318 * NULL is returned if no match was found, but (*parentp) will still 1319 * potentially be adjusted. 1320 * 1321 * This function will also recurse up the chain if the key is not within the 1322 * current parent's range. (*parentp) can never be set to NULL. An iteration 1323 * can simply allow (*parentp) to float inside the loop. 1324 */ 1325 hammer2_chain_t * 1326 hammer2_chain_lookup(hammer2_chain_t **parentp, 1327 hammer2_key_t key_beg, hammer2_key_t key_end, 1328 int flags) 1329 { 1330 hammer2_mount_t *hmp; 1331 hammer2_chain_t *parent; 1332 hammer2_chain_t *chain; 1333 hammer2_chain_t *tmp; 1334 hammer2_blockref_t *base; 1335 hammer2_blockref_t *bref; 1336 hammer2_key_t scan_beg; 1337 hammer2_key_t scan_end; 1338 int count = 0; 1339 int i; 1340 int how_always = HAMMER2_RESOLVE_ALWAYS; 1341 int how_maybe = HAMMER2_RESOLVE_MAYBE; 1342 1343 if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK)) { 1344 how_maybe |= HAMMER2_RESOLVE_SHARED; 1345 how_always |= HAMMER2_RESOLVE_SHARED; 1346 } 1347 1348 /* 1349 * Recurse (*parentp) upward if necessary until the parent completely 1350 * encloses the key range or we hit the inode. 1351 */ 1352 parent = *parentp; 1353 hmp = parent->hmp; 1354 1355 while (parent->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 1356 parent->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 1357 scan_beg = parent->bref.key; 1358 scan_end = scan_beg + 1359 ((hammer2_key_t)1 << parent->bref.keybits) - 1; 1360 if (key_beg >= scan_beg && key_end <= scan_end) 1361 break; 1362 hammer2_chain_ref(parent); /* ref old parent */ 1363 hammer2_chain_unlock(parent); /* unlock old parent */ 1364 parent = parent->parent; 1365 /* lock new parent */ 1366 hammer2_chain_lock(parent, how_maybe); 1367 hammer2_chain_drop(*parentp); /* drop old parent */ 1368 *parentp = parent; /* new parent */ 1369 } 1370 1371 again: 1372 /* 1373 * Locate the blockref array. Currently we do a fully associative 1374 * search through the array. 1375 */ 1376 switch(parent->bref.type) { 1377 case HAMMER2_BREF_TYPE_INODE: 1378 /* 1379 * Special shortcut for embedded data returns the inode 1380 * itself. Callers must detect this condition and access 1381 * the embedded data (the strategy code does this for us). 1382 * 1383 * This is only applicable to regular files and softlinks. 1384 */ 1385 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) { 1386 if (flags & HAMMER2_LOOKUP_NOLOCK) 1387 hammer2_chain_ref(parent); 1388 else 1389 hammer2_chain_lock(parent, how_always); 1390 return (parent); 1391 } 1392 base = &parent->data->ipdata.u.blockset.blockref[0]; 1393 count = HAMMER2_SET_COUNT; 1394 break; 1395 case HAMMER2_BREF_TYPE_INDIRECT: 1396 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1397 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1398 /* 1399 * Optimize indirect blocks in the INITIAL state to avoid 1400 * I/O. 1401 */ 1402 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 1403 base = NULL; 1404 } else { 1405 if (parent->data == NULL) 1406 panic("parent->data is NULL"); 1407 base = &parent->data->npdata.blockref[0]; 1408 } 1409 count = parent->bytes / sizeof(hammer2_blockref_t); 1410 break; 1411 case HAMMER2_BREF_TYPE_VOLUME: 1412 base = &hmp->voldata.sroot_blockset.blockref[0]; 1413 count = HAMMER2_SET_COUNT; 1414 break; 1415 default: 1416 panic("hammer2_chain_lookup: unrecognized blockref type: %d", 1417 parent->bref.type); 1418 base = NULL; /* safety */ 1419 count = 0; /* safety */ 1420 } 1421 1422 /* 1423 * If the element and key overlap we use the element. 1424 * 1425 * NOTE! Deleted elements are effectively invisible. Deletions 1426 * proactively clear the parent bref to the deleted child 1427 * so we do not try to shadow here to avoid parent updates 1428 * (which would be difficult since multiple deleted elements 1429 * might represent different flush synchronization points). 1430 */ 1431 bref = NULL; 1432 for (i = 0; i < count; ++i) { 1433 tmp = hammer2_chain_find(parent, i); 1434 if (tmp) { 1435 KKASSERT((tmp->flags & HAMMER2_CHAIN_DELETED) == 0); 1436 bref = &tmp->bref; 1437 KKASSERT(bref->type != 0); 1438 } else if (base == NULL || base[i].type == 0) { 1439 continue; 1440 } else { 1441 bref = &base[i]; 1442 } 1443 scan_beg = bref->key; 1444 scan_end = scan_beg + ((hammer2_key_t)1 << bref->keybits) - 1; 1445 if (tmp) 1446 hammer2_chain_drop(tmp); 1447 if (key_beg <= scan_end && key_end >= scan_beg) 1448 break; 1449 } 1450 if (i == count) { 1451 if (key_beg == key_end) 1452 return (NULL); 1453 return (hammer2_chain_next(parentp, NULL, 1454 key_beg, key_end, flags)); 1455 } 1456 1457 /* 1458 * Acquire the new chain element. If the chain element is an 1459 * indirect block we must search recursively. 1460 * 1461 * It is possible for the tmp chain above to be removed from 1462 * the RBTREE but the parent lock ensures it would not have been 1463 * destroyed from the media, so the chain_get() code will simply 1464 * reload it from the media in that case. 1465 */ 1466 chain = hammer2_chain_get(parent, i, flags); 1467 if (chain == NULL) 1468 return (NULL); 1469 1470 /* 1471 * If the chain element is an indirect block it becomes the new 1472 * parent and we loop on it. 1473 * 1474 * The parent always has to be locked with at least RESOLVE_MAYBE 1475 * so we can access its data. It might need a fixup if the caller 1476 * passed incompatible flags. Be careful not to cause a deadlock 1477 * as a data-load requires an exclusive lock. 1478 */ 1479 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 1480 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 1481 hammer2_chain_unlock(parent); 1482 *parentp = parent = chain; 1483 if (flags & HAMMER2_LOOKUP_NOLOCK) { 1484 hammer2_chain_lock(chain, how_maybe); 1485 hammer2_chain_drop(chain); /* excess ref */ 1486 } else if ((flags & HAMMER2_LOOKUP_NODATA) && 1487 chain->data == NULL) { 1488 hammer2_chain_ref(chain); 1489 hammer2_chain_unlock(chain); 1490 hammer2_chain_lock(chain, how_maybe | 1491 HAMMER2_RESOLVE_NOREF); 1492 } 1493 goto again; 1494 } 1495 1496 /* 1497 * All done, return the chain 1498 */ 1499 return (chain); 1500 } 1501 1502 /* 1503 * After having issued a lookup we can iterate all matching keys. 1504 * 1505 * If chain is non-NULL we continue the iteration from just after it's index. 1506 * 1507 * If chain is NULL we assume the parent was exhausted and continue the 1508 * iteration at the next parent. 1509 * 1510 * parent must be locked on entry and remains locked throughout. chain's 1511 * lock status must match flags. Chain is always at least referenced. 1512 */ 1513 hammer2_chain_t * 1514 hammer2_chain_next(hammer2_chain_t **parentp, hammer2_chain_t *chain, 1515 hammer2_key_t key_beg, hammer2_key_t key_end, 1516 int flags) 1517 { 1518 hammer2_mount_t *hmp; 1519 hammer2_chain_t *parent; 1520 hammer2_chain_t *tmp; 1521 hammer2_blockref_t *base; 1522 hammer2_blockref_t *bref; 1523 hammer2_key_t scan_beg; 1524 hammer2_key_t scan_end; 1525 int i; 1526 int how_maybe = HAMMER2_RESOLVE_MAYBE; 1527 int count; 1528 1529 if (flags & (HAMMER2_LOOKUP_SHARED | HAMMER2_LOOKUP_NOLOCK)) 1530 how_maybe |= HAMMER2_RESOLVE_SHARED; 1531 1532 parent = *parentp; 1533 hmp = parent->hmp; 1534 1535 again: 1536 /* 1537 * Calculate the next index and recalculate the parent if necessary. 1538 */ 1539 if (chain) { 1540 /* 1541 * Continue iteration within current parent. If not NULL 1542 * the passed-in chain may or may not be locked, based on 1543 * the LOOKUP_NOLOCK flag (passed in as returned from lookup 1544 * or a prior next). 1545 */ 1546 i = chain->index + 1; 1547 if (flags & HAMMER2_LOOKUP_NOLOCK) 1548 hammer2_chain_drop(chain); 1549 else 1550 hammer2_chain_unlock(chain); 1551 1552 /* 1553 * Any scan where the lookup returned degenerate data embedded 1554 * in the inode has an invalid index and must terminate. 1555 */ 1556 if (chain == parent) 1557 return(NULL); 1558 chain = NULL; 1559 } else if (parent->bref.type != HAMMER2_BREF_TYPE_INDIRECT && 1560 parent->bref.type != HAMMER2_BREF_TYPE_FREEMAP_NODE) { 1561 /* 1562 * We reached the end of the iteration. 1563 */ 1564 return (NULL); 1565 } else { 1566 /* 1567 * Continue iteration with next parent unless the current 1568 * parent covers the range. 1569 */ 1570 hammer2_chain_t *nparent; 1571 1572 scan_beg = parent->bref.key; 1573 scan_end = scan_beg + 1574 ((hammer2_key_t)1 << parent->bref.keybits) - 1; 1575 if (key_beg >= scan_beg && key_end <= scan_end) 1576 return (NULL); 1577 1578 i = parent->index + 1; 1579 nparent = parent->parent; 1580 hammer2_chain_ref(nparent); /* ref new parent */ 1581 hammer2_chain_unlock(parent); /* unlock old parent */ 1582 /* lock new parent */ 1583 hammer2_chain_lock(nparent, how_maybe); 1584 hammer2_chain_drop(nparent); /* drop excess ref */ 1585 *parentp = parent = nparent; 1586 } 1587 1588 again2: 1589 /* 1590 * Locate the blockref array. Currently we do a fully associative 1591 * search through the array. 1592 */ 1593 switch(parent->bref.type) { 1594 case HAMMER2_BREF_TYPE_INODE: 1595 base = &parent->data->ipdata.u.blockset.blockref[0]; 1596 count = HAMMER2_SET_COUNT; 1597 break; 1598 case HAMMER2_BREF_TYPE_INDIRECT: 1599 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1600 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1601 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 1602 base = NULL; 1603 } else { 1604 KKASSERT(parent->data != NULL); 1605 base = &parent->data->npdata.blockref[0]; 1606 } 1607 count = parent->bytes / sizeof(hammer2_blockref_t); 1608 break; 1609 case HAMMER2_BREF_TYPE_VOLUME: 1610 base = &hmp->voldata.sroot_blockset.blockref[0]; 1611 count = HAMMER2_SET_COUNT; 1612 break; 1613 default: 1614 panic("hammer2_chain_next: unrecognized blockref type: %d", 1615 parent->bref.type); 1616 base = NULL; /* safety */ 1617 count = 0; /* safety */ 1618 break; 1619 } 1620 KKASSERT(i <= count); 1621 1622 /* 1623 * Look for the key. If we are unable to find a match and an exact 1624 * match was requested we return NULL. If a range was requested we 1625 * run hammer2_chain_next() to iterate. 1626 * 1627 * NOTE! Deleted elements are effectively invisible. Deletions 1628 * proactively clear the parent bref to the deleted child 1629 * so we do not try to shadow here to avoid parent updates 1630 * (which would be difficult since multiple deleted elements 1631 * might represent different flush synchronization points). 1632 */ 1633 bref = NULL; 1634 while (i < count) { 1635 tmp = hammer2_chain_find(parent, i); 1636 if (tmp) { 1637 KKASSERT((tmp->flags & HAMMER2_CHAIN_DELETED) == 0); 1638 bref = &tmp->bref; 1639 } else if (base == NULL || base[i].type == 0) { 1640 ++i; 1641 continue; 1642 } else { 1643 bref = &base[i]; 1644 } 1645 scan_beg = bref->key; 1646 scan_end = scan_beg + ((hammer2_key_t)1 << bref->keybits) - 1; 1647 if (tmp) 1648 hammer2_chain_drop(tmp); 1649 if (key_beg <= scan_end && key_end >= scan_beg) 1650 break; 1651 ++i; 1652 } 1653 1654 /* 1655 * If we couldn't find a match recurse up a parent to continue the 1656 * search. 1657 */ 1658 if (i == count) 1659 goto again; 1660 1661 /* 1662 * Acquire the new chain element. If the chain element is an 1663 * indirect block we must search recursively. 1664 */ 1665 chain = hammer2_chain_get(parent, i, flags); 1666 if (chain == NULL) 1667 return (NULL); 1668 1669 /* 1670 * If the chain element is an indirect block it becomes the new 1671 * parent and we loop on it. 1672 * 1673 * The parent always has to be locked with at least RESOLVE_MAYBE 1674 * so we can access its data. It might need a fixup if the caller 1675 * passed incompatible flags. Be careful not to cause a deadlock 1676 * as a data-load requires an exclusive lock. 1677 */ 1678 if (chain->bref.type == HAMMER2_BREF_TYPE_INDIRECT || 1679 chain->bref.type == HAMMER2_BREF_TYPE_FREEMAP_NODE) { 1680 hammer2_chain_unlock(parent); 1681 *parentp = parent = chain; 1682 chain = NULL; 1683 if (flags & HAMMER2_LOOKUP_NOLOCK) { 1684 hammer2_chain_lock(parent, how_maybe); 1685 hammer2_chain_drop(parent); /* excess ref */ 1686 } else if ((flags & HAMMER2_LOOKUP_NODATA) && 1687 parent->data == NULL) { 1688 hammer2_chain_ref(parent); 1689 hammer2_chain_unlock(parent); 1690 hammer2_chain_lock(parent, how_maybe | 1691 HAMMER2_RESOLVE_NOREF); 1692 } 1693 i = 0; 1694 goto again2; 1695 } 1696 1697 /* 1698 * All done, return chain 1699 */ 1700 return (chain); 1701 } 1702 1703 /* 1704 * Create and return a new hammer2 system memory structure of the specified 1705 * key, type and size and insert it RELATIVE TO (PARENT). 1706 * 1707 * (parent) is typically either an inode or an indirect block, acquired 1708 * acquired as a side effect of issuing a prior failed lookup. parent 1709 * must be locked and held. Do not pass the inode chain to this function 1710 * unless that is the chain returned by the failed lookup. 1711 * 1712 * (chain) is either NULL, a newly allocated chain, or a chain allocated 1713 * via hammer2_chain_duplicate(). When not NULL, the passed-in chain must 1714 * NOT be attached to any parent, and will be attached by this function. 1715 * This mechanic is used by the rename code. 1716 * 1717 * Non-indirect types will automatically allocate indirect blocks as required 1718 * if the new item does not fit in the current (parent). 1719 * 1720 * Indirect types will move a portion of the existing blockref array in 1721 * (parent) into the new indirect type and then use one of the free slots 1722 * to emplace the new indirect type. 1723 * 1724 * A new locked chain element is returned of the specified type. The 1725 * element may or may not have a data area associated with it: 1726 * 1727 * VOLUME not allowed here 1728 * INODE kmalloc()'d data area is set up 1729 * INDIRECT not allowed here 1730 * DATA no data area will be set-up (caller is expected 1731 * to have logical buffers, we don't want to alias 1732 * the data onto device buffers!). 1733 * 1734 * Requires an exclusively locked parent. 1735 */ 1736 int 1737 hammer2_chain_create(hammer2_trans_t *trans, hammer2_chain_t *parent, 1738 hammer2_chain_t **chainp, 1739 hammer2_key_t key, int keybits, int type, size_t bytes) 1740 { 1741 hammer2_mount_t *hmp; 1742 hammer2_chain_t *chain; 1743 hammer2_blockref_t dummy; 1744 hammer2_blockref_t *base; 1745 hammer2_chain_t dummy_chain; 1746 int unlock_parent = 0; 1747 int allocated = 0; 1748 int error = 0; 1749 int count; 1750 int i; 1751 1752 KKASSERT(ccms_thread_lock_owned(&parent->core->cst)); 1753 hmp = parent->hmp; 1754 chain = *chainp; 1755 1756 if (chain == NULL) { 1757 /* 1758 * First allocate media space and construct the dummy bref, 1759 * then allocate the in-memory chain structure. 1760 */ 1761 bzero(&dummy, sizeof(dummy)); 1762 dummy.type = type; 1763 dummy.key = key; 1764 dummy.keybits = keybits; 1765 dummy.data_off = hammer2_allocsize(bytes); 1766 dummy.methods = parent->bref.methods; 1767 chain = hammer2_chain_alloc(hmp, &dummy); 1768 hammer2_chain_core_alloc(chain, NULL); 1769 ccms_thread_lock(&chain->core->cst, CCMS_STATE_EXCLUSIVE); 1770 allocated = 1; 1771 1772 /* 1773 * We do NOT set INITIAL here (yet). INITIAL is only 1774 * used for indirect blocks. 1775 * 1776 * Recalculate bytes to reflect the actual media block 1777 * allocation. 1778 */ 1779 bytes = (hammer2_off_t)1 << 1780 (int)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX); 1781 chain->bytes = bytes; 1782 1783 switch(type) { 1784 case HAMMER2_BREF_TYPE_VOLUME: 1785 panic("hammer2_chain_create: called with volume type"); 1786 break; 1787 case HAMMER2_BREF_TYPE_INODE: 1788 KKASSERT(bytes == HAMMER2_INODE_BYTES); 1789 chain->data = kmalloc(sizeof(chain->data->ipdata), 1790 hmp->minode, M_WAITOK | M_ZERO); 1791 break; 1792 case HAMMER2_BREF_TYPE_INDIRECT: 1793 panic("hammer2_chain_create: cannot be used to" 1794 "create indirect block"); 1795 break; 1796 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1797 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1798 panic("hammer2_chain_create: cannot be used to" 1799 "create freemap root or node"); 1800 break; 1801 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1802 case HAMMER2_BREF_TYPE_DATA: 1803 default: 1804 /* leave chain->data NULL */ 1805 KKASSERT(chain->data == NULL); 1806 break; 1807 } 1808 } else { 1809 /* 1810 * Potentially update the chain's key/keybits. 1811 */ 1812 chain->bref.key = key; 1813 chain->bref.keybits = keybits; 1814 } 1815 1816 again: 1817 /* 1818 * Locate a free blockref in the parent's array 1819 */ 1820 switch(parent->bref.type) { 1821 case HAMMER2_BREF_TYPE_INODE: 1822 KKASSERT((parent->data->ipdata.op_flags & 1823 HAMMER2_OPFLAG_DIRECTDATA) == 0); 1824 KKASSERT(parent->data != NULL); 1825 base = &parent->data->ipdata.u.blockset.blockref[0]; 1826 count = HAMMER2_SET_COUNT; 1827 break; 1828 case HAMMER2_BREF_TYPE_INDIRECT: 1829 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1830 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1831 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 1832 base = NULL; 1833 } else { 1834 KKASSERT(parent->data != NULL); 1835 base = &parent->data->npdata.blockref[0]; 1836 } 1837 count = parent->bytes / sizeof(hammer2_blockref_t); 1838 break; 1839 case HAMMER2_BREF_TYPE_VOLUME: 1840 KKASSERT(parent->data != NULL); 1841 base = &hmp->voldata.sroot_blockset.blockref[0]; 1842 count = HAMMER2_SET_COUNT; 1843 break; 1844 default: 1845 panic("hammer2_chain_create: unrecognized blockref type: %d", 1846 parent->bref.type); 1847 count = 0; 1848 break; 1849 } 1850 1851 /* 1852 * Scan for an unallocated bref, also skipping any slots occupied 1853 * by in-memory chain elements that may not yet have been updated 1854 * in the parent's bref array. 1855 * 1856 * We don't have to hold the spinlock to save an empty slot as 1857 * new slots can only transition from empty if the parent is 1858 * locked exclusively. 1859 */ 1860 bzero(&dummy_chain, sizeof(dummy_chain)); 1861 dummy_chain.delete_tid = HAMMER2_MAX_TID; 1862 1863 spin_lock(&parent->core->cst.spin); 1864 for (i = 0; i < count; ++i) { 1865 if (base == NULL) { 1866 dummy_chain.index = i; 1867 if (RB_FIND(hammer2_chain_tree, 1868 &parent->core->rbtree, &dummy_chain) == NULL) { 1869 break; 1870 } 1871 } else if (base[i].type == 0) { 1872 dummy_chain.index = i; 1873 if (RB_FIND(hammer2_chain_tree, 1874 &parent->core->rbtree, &dummy_chain) == NULL) { 1875 break; 1876 } 1877 } 1878 } 1879 spin_unlock(&parent->core->cst.spin); 1880 1881 /* 1882 * If no free blockref could be found we must create an indirect 1883 * block and move a number of blockrefs into it. With the parent 1884 * locked we can safely lock each child in order to move it without 1885 * causing a deadlock. 1886 * 1887 * This may return the new indirect block or the old parent depending 1888 * on where the key falls. NULL is returned on error. 1889 */ 1890 if (i == count) { 1891 hammer2_chain_t *nparent; 1892 1893 nparent = hammer2_chain_create_indirect(trans, parent, 1894 key, keybits, 1895 &error); 1896 if (nparent == NULL) { 1897 if (allocated) 1898 hammer2_chain_free(chain); 1899 chain = NULL; 1900 goto done; 1901 } 1902 if (parent != nparent) { 1903 if (unlock_parent) 1904 hammer2_chain_unlock(parent); 1905 parent = nparent; 1906 unlock_parent = 1; 1907 } 1908 goto again; 1909 } 1910 1911 /* 1912 * Link the chain into its parent. Later on we will have to set 1913 * the MOVED bit in situations where we don't mark the new chain 1914 * as being modified. 1915 */ 1916 if (chain->parent != NULL) 1917 panic("hammer2: hammer2_chain_create: chain already connected"); 1918 KKASSERT(chain->parent == NULL); 1919 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0); 1920 1921 chain->parent = parent; 1922 chain->index = i; 1923 KKASSERT(parent->refs > 0); 1924 spin_lock(&parent->core->cst.spin); 1925 if (RB_INSERT(hammer2_chain_tree, &parent->core->rbtree, chain)) 1926 panic("hammer2_chain_link: collision"); 1927 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 1928 hammer2_chain_ref(parent); /* chain->parent ref */ 1929 spin_unlock(&parent->core->cst.spin); 1930 1931 /* 1932 * (allocated) indicates that this is a newly-created chain element 1933 * rather than a renamed chain element. In this situation we want 1934 * to place the chain element in the MODIFIED state. 1935 * 1936 * The data area will be set up as follows: 1937 * 1938 * VOLUME not allowed here. 1939 * 1940 * INODE embedded data are will be set-up. 1941 * 1942 * INDIRECT not allowed here. 1943 * 1944 * DATA no data area will be set-up (caller is expected 1945 * to have logical buffers, we don't want to alias 1946 * the data onto device buffers!). 1947 */ 1948 if (allocated) { 1949 switch(chain->bref.type) { 1950 case HAMMER2_BREF_TYPE_DATA: 1951 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 1952 hammer2_chain_modify(trans, chain, 1953 HAMMER2_MODIFY_OPTDATA); 1954 break; 1955 case HAMMER2_BREF_TYPE_INDIRECT: 1956 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 1957 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 1958 /* not supported in this function */ 1959 panic("hammer2_chain_create: bad type"); 1960 atomic_set_int(&chain->flags, HAMMER2_CHAIN_INITIAL); 1961 hammer2_chain_modify(trans, chain, 1962 HAMMER2_MODIFY_OPTDATA); 1963 break; 1964 default: 1965 hammer2_chain_modify(trans, chain, 0); 1966 break; 1967 } 1968 } else { 1969 /* 1970 * When reconnecting inodes we have to call setsubmod() 1971 * to ensure that its state propagates up the newly 1972 * connected parent. 1973 * 1974 * Make sure MOVED is set but do not update bref_flush. If 1975 * the chain is undergoing modification bref_flush will be 1976 * updated when it gets flushed. If it is not then the 1977 * bref may not have been flushed yet and we do not want to 1978 * set MODIFIED here as this could result in unnecessary 1979 * reallocations. 1980 */ 1981 if ((chain->flags & HAMMER2_CHAIN_MOVED) == 0) { 1982 hammer2_chain_ref(chain); 1983 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MOVED); 1984 } 1985 hammer2_chain_parent_setsubmod(chain); 1986 } 1987 1988 done: 1989 *chainp = chain; 1990 if (unlock_parent) 1991 hammer2_chain_unlock(parent); 1992 return (error); 1993 } 1994 1995 /* 1996 * Replace (*chainp) with a duplicate. The original *chainp is unlocked 1997 * and the replacement will be returned locked. Both the original and the 1998 * new chain will share the same RBTREE (have the same chain->core), with 1999 * the new chain becoming the 'current' chain (meaning it is the first in 2000 * the linked list at core->chain_first). 2001 * 2002 * If (parent, i) then the new duplicated chain is inserted under the parent 2003 * at the specified index (the parent must not have a ref at that index). 2004 * 2005 * If (NULL, -1) then the new duplicated chain is not inserted anywhere, 2006 * similar to if it had just been chain_alloc()'d (suitable for passing into 2007 * hammer2_chain_create() after this function returns). 2008 * 2009 * NOTE! Duplication is used in order to retain the original topology to 2010 * support flush synchronization points. Both the original and the 2011 * new chain will have the same transaction id and thus the operation 2012 * appears atomic on the media. 2013 */ 2014 void 2015 hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t *parent, 2016 int i, hammer2_chain_t **chainp) 2017 { 2018 hammer2_mount_t *hmp = trans->hmp; 2019 hammer2_blockref_t *base; 2020 hammer2_chain_t *chain; 2021 size_t bytes; 2022 int count; 2023 2024 /* 2025 * First create a duplicate of the chain structure, associating 2026 * it with the same core, making it the same size, pointing it 2027 * to the same bref (the same media block), and copying any inline 2028 * data. 2029 */ 2030 KKASSERT(((*chainp)->flags & HAMMER2_CHAIN_INITIAL) == 0); 2031 chain = hammer2_chain_alloc(hmp, &(*chainp)->bref); 2032 hammer2_chain_core_alloc(chain, (*chainp)->core); 2033 2034 bytes = (hammer2_off_t)1 << 2035 (int)(chain->bref.data_off & HAMMER2_OFF_MASK_RADIX); 2036 chain->bytes = bytes; 2037 2038 switch(chain->bref.type) { 2039 case HAMMER2_BREF_TYPE_VOLUME: 2040 panic("hammer2_chain_duplicate: cannot be called w/volhdr"); 2041 break; 2042 case HAMMER2_BREF_TYPE_INODE: 2043 KKASSERT(bytes == HAMMER2_INODE_BYTES); 2044 if ((*chainp)->data) { 2045 chain->data = kmalloc(sizeof(chain->data->ipdata), 2046 hmp->minode, M_WAITOK | M_ZERO); 2047 chain->data->ipdata = (*chainp)->data->ipdata; 2048 } 2049 break; 2050 case HAMMER2_BREF_TYPE_INDIRECT: 2051 #if 0 2052 panic("hammer2_chain_duplicate: cannot be used to" 2053 "create an indirect block"); 2054 #endif 2055 break; 2056 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2057 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2058 panic("hammer2_chain_duplicate: cannot be used to" 2059 "create a freemap root or node"); 2060 break; 2061 case HAMMER2_BREF_TYPE_FREEMAP_LEAF: 2062 case HAMMER2_BREF_TYPE_DATA: 2063 default: 2064 /* leave chain->data NULL */ 2065 KKASSERT(chain->data == NULL); 2066 break; 2067 } 2068 2069 /* 2070 * Both chains must be locked for us to be able to set the 2071 * duplink. To avoid buffer cache deadlocks we do not try 2072 * to resolve the new chain until after we've unlocked the 2073 * old one. 2074 */ 2075 hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER); 2076 KKASSERT((*chainp)->duplink == NULL); 2077 (*chainp)->duplink = chain; /* inherits excess ref from alloc */ 2078 hammer2_chain_unlock(*chainp); 2079 *chainp = chain; 2080 hammer2_chain_lock(chain, HAMMER2_RESOLVE_MAYBE); 2081 hammer2_chain_unlock(chain); 2082 2083 2084 /* 2085 * If parent is not NULL, insert into the parent at the requested 2086 * index. The newly duplicated chain must be marked MOVED and 2087 * SUBMODIFIED set in its parent(s). 2088 */ 2089 if (parent) { 2090 /* 2091 * Locate a free blockref in the parent's array 2092 */ 2093 KKASSERT(ccms_thread_lock_owned(&parent->core->cst)); 2094 switch(parent->bref.type) { 2095 case HAMMER2_BREF_TYPE_INODE: 2096 KKASSERT((parent->data->ipdata.op_flags & 2097 HAMMER2_OPFLAG_DIRECTDATA) == 0); 2098 KKASSERT(parent->data != NULL); 2099 base = &parent->data->ipdata.u.blockset.blockref[0]; 2100 count = HAMMER2_SET_COUNT; 2101 break; 2102 case HAMMER2_BREF_TYPE_INDIRECT: 2103 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2104 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2105 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 2106 base = NULL; 2107 } else { 2108 KKASSERT(parent->data != NULL); 2109 base = &parent->data->npdata.blockref[0]; 2110 } 2111 count = parent->bytes / sizeof(hammer2_blockref_t); 2112 break; 2113 case HAMMER2_BREF_TYPE_VOLUME: 2114 KKASSERT(parent->data != NULL); 2115 base = &hmp->voldata.sroot_blockset.blockref[0]; 2116 count = HAMMER2_SET_COUNT; 2117 break; 2118 default: 2119 panic("hammer2_chain_create: unrecognized " 2120 "blockref type: %d", 2121 parent->bref.type); 2122 count = 0; 2123 break; 2124 } 2125 KKASSERT(i >= 0 && i < count); 2126 KKASSERT(base == NULL || base[i].type == 0); 2127 2128 chain->parent = parent; 2129 chain->index = i; 2130 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0); 2131 KKASSERT(parent->refs > 0); 2132 spin_lock(&parent->core->cst.spin); 2133 if (RB_INSERT(hammer2_chain_tree, &parent->core->rbtree, chain)) 2134 panic("hammer2_chain_link: collision"); 2135 atomic_set_int(&chain->flags, HAMMER2_CHAIN_ONRBTREE); 2136 hammer2_chain_ref(parent); /* chain->parent ref */ 2137 spin_unlock(&parent->core->cst.spin); 2138 2139 if ((chain->flags & HAMMER2_CHAIN_MOVED) == 0) { 2140 hammer2_chain_ref(chain); 2141 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MOVED); 2142 } 2143 hammer2_chain_parent_setsubmod(chain); 2144 } 2145 } 2146 2147 /* 2148 * Create an indirect block that covers one or more of the elements in the 2149 * current parent. Either returns the existing parent with no locking or 2150 * ref changes or returns the new indirect block locked and referenced 2151 * and leaving the original parent lock/ref intact as well. 2152 * 2153 * If an error occurs, NULL is returned and *errorp is set to the error. 2154 * 2155 * The returned chain depends on where the specified key falls. 2156 * 2157 * The key/keybits for the indirect mode only needs to follow three rules: 2158 * 2159 * (1) That all elements underneath it fit within its key space and 2160 * 2161 * (2) That all elements outside it are outside its key space. 2162 * 2163 * (3) When creating the new indirect block any elements in the current 2164 * parent that fit within the new indirect block's keyspace must be 2165 * moved into the new indirect block. 2166 * 2167 * (4) The keyspace chosen for the inserted indirect block CAN cover a wider 2168 * keyspace the the current parent, but lookup/iteration rules will 2169 * ensure (and must ensure) that rule (2) for all parents leading up 2170 * to the nearest inode or the root volume header is adhered to. This 2171 * is accomplished by always recursing through matching keyspaces in 2172 * the hammer2_chain_lookup() and hammer2_chain_next() API. 2173 * 2174 * The current implementation calculates the current worst-case keyspace by 2175 * iterating the current parent and then divides it into two halves, choosing 2176 * whichever half has the most elements (not necessarily the half containing 2177 * the requested key). 2178 * 2179 * We can also opt to use the half with the least number of elements. This 2180 * causes lower-numbered keys (aka logical file offsets) to recurse through 2181 * fewer indirect blocks and higher-numbered keys to recurse through more. 2182 * This also has the risk of not moving enough elements to the new indirect 2183 * block and being forced to create several indirect blocks before the element 2184 * can be inserted. 2185 * 2186 * Must be called with an exclusively locked parent. 2187 */ 2188 static 2189 hammer2_chain_t * 2190 hammer2_chain_create_indirect(hammer2_trans_t *trans, hammer2_chain_t *parent, 2191 hammer2_key_t create_key, int create_bits, 2192 int *errorp) 2193 { 2194 hammer2_mount_t *hmp = trans->hmp; 2195 hammer2_blockref_t *base; 2196 hammer2_blockref_t *bref; 2197 hammer2_chain_t *chain; 2198 hammer2_chain_t *ichain; 2199 hammer2_chain_t dummy; 2200 hammer2_key_t key = create_key; 2201 int keybits = create_bits; 2202 int locount = 0; 2203 int hicount = 0; 2204 int count; 2205 int nbytes; 2206 int i; 2207 2208 /* 2209 * Calculate the base blockref pointer or NULL if the chain 2210 * is known to be empty. We need to calculate the array count 2211 * for RB lookups either way. 2212 */ 2213 KKASSERT(ccms_thread_lock_owned(&parent->core->cst)); 2214 *errorp = 0; 2215 2216 hammer2_chain_modify(trans, parent, HAMMER2_MODIFY_OPTDATA); 2217 if (parent->flags & HAMMER2_CHAIN_INITIAL) { 2218 base = NULL; 2219 2220 switch(parent->bref.type) { 2221 case HAMMER2_BREF_TYPE_INODE: 2222 count = HAMMER2_SET_COUNT; 2223 break; 2224 case HAMMER2_BREF_TYPE_INDIRECT: 2225 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2226 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2227 count = parent->bytes / sizeof(hammer2_blockref_t); 2228 break; 2229 case HAMMER2_BREF_TYPE_VOLUME: 2230 count = HAMMER2_SET_COUNT; 2231 break; 2232 default: 2233 panic("hammer2_chain_create_indirect: " 2234 "unrecognized blockref type: %d", 2235 parent->bref.type); 2236 count = 0; 2237 break; 2238 } 2239 } else { 2240 switch(parent->bref.type) { 2241 case HAMMER2_BREF_TYPE_INODE: 2242 base = &parent->data->ipdata.u.blockset.blockref[0]; 2243 count = HAMMER2_SET_COUNT; 2244 break; 2245 case HAMMER2_BREF_TYPE_INDIRECT: 2246 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2247 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2248 base = &parent->data->npdata.blockref[0]; 2249 count = parent->bytes / sizeof(hammer2_blockref_t); 2250 break; 2251 case HAMMER2_BREF_TYPE_VOLUME: 2252 base = &hmp->voldata.sroot_blockset.blockref[0]; 2253 count = HAMMER2_SET_COUNT; 2254 break; 2255 default: 2256 panic("hammer2_chain_create_indirect: " 2257 "unrecognized blockref type: %d", 2258 parent->bref.type); 2259 count = 0; 2260 break; 2261 } 2262 } 2263 2264 /* 2265 * Scan for an unallocated bref, also skipping any slots occupied 2266 * by in-memory chain elements which may not yet have been updated 2267 * in the parent's bref array. 2268 * 2269 * Deleted elements are ignored. 2270 */ 2271 bzero(&dummy, sizeof(dummy)); 2272 dummy.delete_tid = HAMMER2_MAX_TID; 2273 2274 spin_lock(&parent->core->cst.spin); 2275 for (i = 0; i < count; ++i) { 2276 int nkeybits; 2277 2278 dummy.index = i; 2279 chain = RB_FIND(hammer2_chain_tree, &parent->core->rbtree, 2280 &dummy); 2281 if (chain) { 2282 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0); 2283 bref = &chain->bref; 2284 } else if (base && base[i].type) { 2285 bref = &base[i]; 2286 } else { 2287 continue; 2288 } 2289 2290 /* 2291 * Expand our calculated key range (key, keybits) to fit 2292 * the scanned key. nkeybits represents the full range 2293 * that we will later cut in half (two halves @ nkeybits - 1). 2294 */ 2295 nkeybits = keybits; 2296 if (nkeybits < bref->keybits) 2297 nkeybits = bref->keybits; 2298 while (nkeybits < 64 && 2299 (~(((hammer2_key_t)1 << nkeybits) - 1) & 2300 (key ^ bref->key)) != 0) { 2301 ++nkeybits; 2302 } 2303 2304 /* 2305 * If the new key range is larger we have to determine 2306 * which side of the new key range the existing keys fall 2307 * under by checking the high bit, then collapsing the 2308 * locount into the hicount or vise-versa. 2309 */ 2310 if (keybits != nkeybits) { 2311 if (((hammer2_key_t)1 << (nkeybits - 1)) & key) { 2312 hicount += locount; 2313 locount = 0; 2314 } else { 2315 locount += hicount; 2316 hicount = 0; 2317 } 2318 keybits = nkeybits; 2319 } 2320 2321 /* 2322 * The newly scanned key will be in the lower half or the 2323 * higher half of the (new) key range. 2324 */ 2325 if (((hammer2_key_t)1 << (nkeybits - 1)) & bref->key) 2326 ++hicount; 2327 else 2328 ++locount; 2329 } 2330 spin_unlock(&parent->core->cst.spin); 2331 bref = NULL; /* now invalid (safety) */ 2332 2333 /* 2334 * Adjust keybits to represent half of the full range calculated 2335 * above (radix 63 max) 2336 */ 2337 --keybits; 2338 2339 /* 2340 * Select whichever half contains the most elements. Theoretically 2341 * we can select either side as long as it contains at least one 2342 * element (in order to ensure that a free slot is present to hold 2343 * the indirect block). 2344 */ 2345 key &= ~(((hammer2_key_t)1 << keybits) - 1); 2346 if (hammer2_indirect_optimize) { 2347 /* 2348 * Insert node for least number of keys, this will arrange 2349 * the first few blocks of a large file or the first few 2350 * inodes in a directory with fewer indirect blocks when 2351 * created linearly. 2352 */ 2353 if (hicount < locount && hicount != 0) 2354 key |= (hammer2_key_t)1 << keybits; 2355 else 2356 key &= ~(hammer2_key_t)1 << keybits; 2357 } else { 2358 /* 2359 * Insert node for most number of keys, best for heavily 2360 * fragmented files. 2361 */ 2362 if (hicount > locount) 2363 key |= (hammer2_key_t)1 << keybits; 2364 else 2365 key &= ~(hammer2_key_t)1 << keybits; 2366 } 2367 2368 /* 2369 * How big should our new indirect block be? It has to be at least 2370 * as large as its parent. 2371 */ 2372 if (parent->bref.type == HAMMER2_BREF_TYPE_INODE) 2373 nbytes = HAMMER2_IND_BYTES_MIN; 2374 else 2375 nbytes = HAMMER2_IND_BYTES_MAX; 2376 if (nbytes < count * sizeof(hammer2_blockref_t)) 2377 nbytes = count * sizeof(hammer2_blockref_t); 2378 2379 /* 2380 * Ok, create our new indirect block 2381 */ 2382 switch(parent->bref.type) { 2383 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2384 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2385 dummy.bref.type = HAMMER2_BREF_TYPE_FREEMAP_NODE; 2386 break; 2387 default: 2388 dummy.bref.type = HAMMER2_BREF_TYPE_INDIRECT; 2389 break; 2390 } 2391 dummy.bref.key = key; 2392 dummy.bref.keybits = keybits; 2393 dummy.bref.data_off = hammer2_allocsize(nbytes); 2394 dummy.bref.methods = parent->bref.methods; 2395 2396 ichain = hammer2_chain_alloc(hmp, &dummy.bref); 2397 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_INITIAL); 2398 hammer2_chain_core_alloc(ichain, NULL); 2399 hammer2_chain_lock(ichain, HAMMER2_RESOLVE_MAYBE); 2400 hammer2_chain_drop(ichain); /* excess ref from alloc */ 2401 2402 /* 2403 * Iterate the original parent and move the matching brefs into 2404 * the new indirect block. 2405 * 2406 * XXX handle flushes. 2407 */ 2408 spin_lock(&parent->core->cst.spin); 2409 for (i = 0; i < count; ++i) { 2410 /* 2411 * For keying purposes access the bref from the media or 2412 * from our in-memory cache. In cases where the in-memory 2413 * cache overrides the media the keyrefs will be the same 2414 * anyway so we can avoid checking the cache when the media 2415 * has a key. 2416 */ 2417 dummy.index = i; 2418 chain = RB_FIND(hammer2_chain_tree, &parent->core->rbtree, 2419 &dummy); 2420 if (chain) { 2421 KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0); 2422 bref = &chain->bref; 2423 } else if (base && base[i].type) { 2424 bref = &base[i]; 2425 } else { 2426 if (ichain->index < 0) 2427 ichain->index = i; 2428 continue; 2429 } 2430 2431 /* 2432 * Skip keys not in the chosen half (low or high), only bit 2433 * (keybits - 1) needs to be compared but for safety we 2434 * will compare all msb bits plus that bit again. 2435 */ 2436 if ((~(((hammer2_key_t)1 << keybits) - 1) & 2437 (key ^ bref->key)) != 0) { 2438 continue; 2439 } 2440 2441 /* 2442 * This element is being moved from the parent, its slot 2443 * is available for our new indirect block. 2444 */ 2445 if (ichain->index < 0) 2446 ichain->index = i; 2447 2448 /* 2449 * Load the new indirect block by acquiring or allocating 2450 * the related chain entries, then move them to the new 2451 * parent (ichain) by deleting them from their old location 2452 * and inserting a duplicate of the chain and any modified 2453 * sub-chain in the new location. 2454 * 2455 * We must set MOVED in the chain being duplicated and 2456 * SUBMODIFIED in the parent(s) so the flush code knows 2457 * what is going on. The latter is done after the loop. 2458 * 2459 * WARNING! chain->cst.spin must be held when chain->parent is 2460 * modified, even though we own the full blown lock, 2461 * to deal with setsubmod and rename races. 2462 * (XXX remove this req). 2463 */ 2464 spin_unlock(&parent->core->cst.spin); 2465 chain = hammer2_chain_get(parent, i, HAMMER2_LOOKUP_NODATA); 2466 hammer2_chain_delete(trans, parent, chain); 2467 hammer2_chain_duplicate(trans, ichain, i, &chain); 2468 2469 #if 0 2470 if (base) 2471 bzero(&base[i], sizeof(base[i])); 2472 #endif 2473 hammer2_chain_unlock(chain); 2474 KKASSERT(parent->refs > 0); 2475 chain = NULL; 2476 spin_lock(&parent->core->cst.spin); 2477 } 2478 spin_unlock(&parent->core->cst.spin); 2479 2480 /* 2481 * Insert the new indirect block into the parent now that we've 2482 * cleared out some entries in the parent. We calculated a good 2483 * insertion index in the loop above (ichain->index). 2484 * 2485 * We don't have to set MOVED here because we mark ichain modified 2486 * down below (so the normal modified -> flush -> set-moved sequence 2487 * applies). 2488 * 2489 * The insertion shouldn't race as this is a completely new block 2490 * and the parent is locked. 2491 */ 2492 KKASSERT(ichain->index >= 0); 2493 KKASSERT((ichain->flags & HAMMER2_CHAIN_ONRBTREE) == 0); 2494 spin_lock(&parent->core->cst.spin); 2495 if (RB_INSERT(hammer2_chain_tree, &parent->core->rbtree, ichain)) 2496 panic("hammer2_chain_create_indirect: ichain insertion"); 2497 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_ONRBTREE); 2498 ichain->parent = parent; 2499 hammer2_chain_ref(parent); /* ichain->parent ref */ 2500 spin_unlock(&parent->core->cst.spin); 2501 2502 /* 2503 * Mark the new indirect block modified after insertion, which 2504 * will propagate up through parent all the way to the root and 2505 * also allocate the physical block in ichain for our caller, 2506 * and assign ichain->data to a pre-zero'd space (because there 2507 * is not prior data to copy into it). 2508 * 2509 * We have to set SUBMODIFIED in ichain's flags manually so the 2510 * flusher knows it has to recurse through it to get to all of 2511 * our moved blocks, then call setsubmod() to set the bit 2512 * recursively. 2513 */ 2514 hammer2_chain_modify(trans, ichain, HAMMER2_MODIFY_OPTDATA); 2515 hammer2_chain_parent_setsubmod(ichain); 2516 atomic_set_int(&ichain->flags, HAMMER2_CHAIN_SUBMODIFIED); 2517 2518 /* 2519 * Figure out what to return. 2520 */ 2521 if (create_bits > keybits) { 2522 /* 2523 * Key being created is way outside the key range, 2524 * return the original parent. 2525 */ 2526 hammer2_chain_unlock(ichain); 2527 } else if (~(((hammer2_key_t)1 << keybits) - 1) & 2528 (create_key ^ key)) { 2529 /* 2530 * Key being created is outside the key range, 2531 * return the original parent. 2532 */ 2533 hammer2_chain_unlock(ichain); 2534 } else { 2535 /* 2536 * Otherwise its in the range, return the new parent. 2537 * (leave both the new and old parent locked). 2538 */ 2539 parent = ichain; 2540 } 2541 2542 return(parent); 2543 } 2544 2545 /* 2546 * Sets CHAIN_DELETED and CHAIN_MOVED in the chain being deleted and 2547 * remove the parent's bref reference to chain, generating a modification 2548 * on the parent. 2549 * 2550 * We do not attempt to defer adjustment of the parent bref to the chain 2551 * as this could become quite complex with multiple deletions / replacements. 2552 * Intead, a modification is generated in the parent which can cause it to 2553 * be duplicated if the current parent's data is required for a flush in 2554 * progress. 2555 * 2556 * NOTE: We can trivially adjust the parent if it is in the INITIAL state. 2557 * 2558 * NOTE: The flush code handles the actual removal of the chain from 2559 * the BTREE (also, depending on synchronization points, the 2560 * chain may still be relevant to the flush). 2561 * 2562 * NOTE: chain->delete_tid distinguishes deleted chains from live chains, 2563 * by setting it to something less than HAMMER2_MAX_TID the 2564 * chain_lookup(), chain_next(), and chain_get() functions will 2565 * not have visibility. 2566 * 2567 * This function is NOT recursive. Any entity already pushed into the 2568 * chain (such as an inode) may still need visibility into its contents, 2569 * as well as the ability to read and modify the contents. For example, 2570 * for an unlinked file which is still open. 2571 */ 2572 void 2573 hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *parent, 2574 hammer2_chain_t *chain) 2575 { 2576 hammer2_mount_t *hmp = trans->hmp; 2577 hammer2_blockref_t *base; 2578 int count; 2579 2580 if (chain->parent != parent) 2581 panic("hammer2_chain_delete: parent mismatch"); 2582 KKASSERT(ccms_thread_lock_owned(&parent->core->cst)); 2583 2584 /* 2585 * Nothing to do if already marked. 2586 */ 2587 if (chain->flags & HAMMER2_CHAIN_DELETED) 2588 return; 2589 2590 /* 2591 * Mark the parent modified so our base[] pointer remains valid 2592 * while we move entries. For the optimized indirect block 2593 * case mark the parent moved instead. 2594 * 2595 * Calculate the blockref reference in the parent and zero it out. 2596 */ 2597 switch(parent->bref.type) { 2598 case HAMMER2_BREF_TYPE_INODE: 2599 hammer2_chain_modify(trans, parent, 2600 HAMMER2_MODIFY_NO_MODIFY_TID); 2601 base = &parent->data->ipdata.u.blockset.blockref[0]; 2602 count = HAMMER2_SET_COUNT; 2603 break; 2604 case HAMMER2_BREF_TYPE_INDIRECT: 2605 case HAMMER2_BREF_TYPE_FREEMAP_ROOT: 2606 case HAMMER2_BREF_TYPE_FREEMAP_NODE: 2607 hammer2_chain_modify(trans, parent, 2608 HAMMER2_MODIFY_OPTDATA | 2609 HAMMER2_MODIFY_NO_MODIFY_TID); 2610 if (parent->flags & HAMMER2_CHAIN_INITIAL) 2611 base = NULL; 2612 else 2613 base = &parent->data->npdata.blockref[0]; 2614 count = parent->bytes / sizeof(hammer2_blockref_t); 2615 break; 2616 case HAMMER2_BREF_TYPE_VOLUME: 2617 hammer2_chain_modify(trans, parent, 2618 HAMMER2_MODIFY_NO_MODIFY_TID); 2619 base = &hmp->voldata.sroot_blockset.blockref[0]; 2620 count = HAMMER2_SET_COUNT; 2621 break; 2622 default: 2623 panic("hammer2_chain_delete: unrecognized blockref type: %d", 2624 parent->bref.type); 2625 base = NULL; /* NOT REACHED */ 2626 count = 0; /* NOT REACHED */ 2627 break; /* NOT REACHED */ 2628 } 2629 KKASSERT(chain->index >= 0 && chain->index < count); 2630 2631 /* 2632 * Clean out the blockref immediately. 2633 */ 2634 if (base) 2635 bzero(&base[chain->index], sizeof(*base)); 2636 2637 /* 2638 * Must set MOVED along with DELETED for the flush code to recognize 2639 * the operation and properly disconnect the chain in-memory. 2640 * 2641 * The setting of DELETED causes finds, lookups, and _next iterations 2642 * to no longer recognize the chain. RB_SCAN()s will still have 2643 * visibility (needed for flush serialization points). 2644 */ 2645 atomic_set_int(&chain->flags, HAMMER2_CHAIN_DELETED); 2646 if ((chain->flags & HAMMER2_CHAIN_MOVED) == 0) { 2647 hammer2_chain_ref(chain); 2648 atomic_set_int(&chain->flags, HAMMER2_CHAIN_MOVED); 2649 } 2650 chain->delete_tid = trans->sync_tid; 2651 hammer2_chain_parent_setsubmod(chain); 2652 } 2653 2654 void 2655 hammer2_chain_wait(hammer2_chain_t *chain) 2656 { 2657 tsleep(chain, 0, "chnflw", 1); 2658 } 2659