1 /* 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 2022 Tomohiro Kusumi <tkusumi@netbsd.org> 5 * Copyright (c) 2011-2022 The DragonFly Project. All rights reserved. 6 * 7 * This code is derived from software contributed to The DragonFly Project 8 * by Matthew Dillon <dillon@dragonflybsd.org> 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * 1. Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 2. Redistributions in binary form must reproduce the above copyright 17 * notice, this list of conditions and the following disclaimer in 18 * the documentation and/or other materials provided with the 19 * distribution. 20 * 3. Neither the name of The DragonFly Project nor the names of its 21 * contributors may be used to endorse or promote products derived 22 * from this software without specific, prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 */ 37 /* 38 #include <sys/cdefs.h> 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/types.h> 42 #include <sys/lock.h> 43 #include <sys/uuid.h> 44 #include <sys/vnode.h> 45 */ 46 47 #include "hammer2.h" 48 49 #define INODE_DEBUG 0 50 51 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp, 52 hammer2_tid_t, meta.inum); 53 54 int 55 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2) 56 { 57 if (ip1->meta.inum < ip2->meta.inum) 58 return(-1); 59 if (ip1->meta.inum > ip2->meta.inum) 60 return(1); 61 return(0); 62 } 63 64 /* 65 * Caller holds pmp->list_spin and the inode should be locked. Merge ip 66 * with the specified depend. 67 * 68 * If the ip is on SYNCQ it stays there and (void *)-1 is returned, indicating 69 * that successive calls must ensure the ip is on a pass2 depend (or they are 70 * all SYNCQ). If the passed-in depend is not NULL and not (void *)-1 then 71 * we can set pass2 on it and return. 72 * 73 * If the ip is not on SYNCQ it is merged with the passed-in depend, creating 74 * a self-depend if necessary, and depend->pass2 is set according 75 * to the PASS2 flag. SIDEQ is set. 76 */ 77 static __noinline 78 hammer2_depend_t * 79 hammer2_inode_setdepend_locked(hammer2_inode_t *ip, hammer2_depend_t *depend) 80 { 81 hammer2_pfs_t *pmp = ip->pmp; 82 hammer2_depend_t *dtmp; 83 hammer2_inode_t *iptmp; 84 85 /* 86 * If ip is SYNCQ its entry is used for the syncq list and it will 87 * no longer be associated with a dependency. Merging this status 88 * with a passed-in depend implies PASS2. 89 */ 90 if (ip->flags & HAMMER2_INODE_SYNCQ) { 91 if (depend == (void *)-1 || 92 depend == NULL) { 93 return ((void *)-1); 94 } 95 depend->pass2 = 1; 96 hammer2_trans_setflags(pmp, HAMMER2_TRANS_RESCAN); 97 98 return depend; 99 } 100 101 /* 102 * If ip is already SIDEQ, merge ip->depend into the passed-in depend. 103 * If it is not, associate the ip with the passed-in depend, creating 104 * a single-entry dependency using depend_static if necessary. 105 * 106 * NOTE: The use of ip->depend_static always requires that the 107 * specific ip containing the structure is part of that 108 * particular depend_static's dependency group. 109 */ 110 if (ip->flags & HAMMER2_INODE_SIDEQ) { 111 /* 112 * Merge ip->depend with the passed-in depend. If the 113 * passed-in depend is not a special case, all ips associated 114 * with ip->depend (including the original ip) must be moved 115 * to the passed-in depend. 116 */ 117 if (depend == NULL) { 118 depend = ip->depend; 119 } else if (depend == (void *)-1) { 120 depend = ip->depend; 121 depend->pass2 = 1; 122 } else if (depend != ip->depend) { 123 #ifdef INVARIANTS 124 int sanitychk = 0; 125 #endif 126 dtmp = ip->depend; 127 while ((iptmp = TAILQ_FIRST(&dtmp->sideq)) != NULL) { 128 #ifdef INVARIANTS 129 if (iptmp == ip) 130 sanitychk = 1; 131 #endif 132 TAILQ_REMOVE(&dtmp->sideq, iptmp, entry); 133 TAILQ_INSERT_TAIL(&depend->sideq, iptmp, entry); 134 iptmp->depend = depend; 135 } 136 KKASSERT(sanitychk == 1); 137 depend->count += dtmp->count; 138 depend->pass2 |= dtmp->pass2; 139 TAILQ_REMOVE(&pmp->depq, dtmp, entry); 140 dtmp->count = 0; 141 dtmp->pass2 = 0; 142 } 143 } else { 144 /* 145 * Add ip to the sideq, creating a self-dependency if 146 * necessary. 147 */ 148 hammer2_inode_ref(ip); /* extra ref usually via hammer2_inode_modify() */ 149 atomic_set_int(&ip->flags, HAMMER2_INODE_SIDEQ); 150 if (depend == NULL) { 151 depend = &ip->depend_static; 152 TAILQ_INSERT_TAIL(&pmp->depq, depend, entry); 153 } else if (depend == (void *)-1) { 154 depend = &ip->depend_static; 155 depend->pass2 = 1; 156 TAILQ_INSERT_TAIL(&pmp->depq, depend, entry); 157 } /* else add ip to passed-in depend */ 158 TAILQ_INSERT_TAIL(&depend->sideq, ip, entry); 159 ip->depend = depend; 160 ++depend->count; 161 ++pmp->sideq_count; 162 } 163 164 if (ip->flags & HAMMER2_INODE_SYNCQ_PASS2) 165 depend->pass2 = 1; 166 if (depend->pass2) 167 hammer2_trans_setflags(pmp, HAMMER2_TRANS_RESCAN); 168 169 return depend; 170 } 171 172 /* 173 * Put a solo inode on the SIDEQ (meaning that its dirty). This can also 174 * occur from inode_lock4() and inode_depend(). 175 * 176 * Caller must pass-in a locked inode. 177 */ 178 void 179 hammer2_inode_delayed_sideq(hammer2_inode_t *ip) 180 { 181 hammer2_pfs_t *pmp = ip->pmp; 182 183 /* 184 * Optimize case to avoid pmp spinlock. 185 */ 186 if ((ip->flags & (HAMMER2_INODE_SYNCQ | HAMMER2_INODE_SIDEQ)) == 0) { 187 hammer2_spin_ex(&pmp->list_spin); 188 hammer2_inode_setdepend_locked(ip, NULL); 189 hammer2_spin_unex(&pmp->list_spin); 190 } 191 } 192 193 /* 194 * Lock an inode, with SYNCQ semantics. 195 * 196 * HAMMER2 offers shared and exclusive locks on inodes. Pass a mask of 197 * flags for options: 198 * 199 * - pass HAMMER2_RESOLVE_SHARED if a shared lock is desired. 200 * shared locks are not subject to SYNCQ semantics, exclusive locks 201 * are. 202 * 203 * - pass HAMMER2_RESOLVE_ALWAYS if you need the inode's meta-data. 204 * Most front-end inode locks do. 205 * 206 * - pass HAMMER2_RESOLVE_NEVER if you do not want to require that 207 * the inode data be resolved. This is used by the syncthr because 208 * it can run on an unresolved/out-of-sync cluster, and also by the 209 * vnode reclamation code to avoid unnecessary I/O (particularly when 210 * disposing of hundreds of thousands of cached vnodes). 211 * 212 * This function, along with lock4, has SYNCQ semantics. If the inode being 213 * locked is on the SYNCQ, that is it has been staged by the syncer, we must 214 * block until the operation is complete (even if we can lock the inode). In 215 * order to reduce the stall time, we re-order the inode to the front of the 216 * pmp->syncq prior to blocking. This reordering VERY significantly improves 217 * performance. 218 * 219 * The inode locking function locks the inode itself, resolves any stale 220 * chains in the inode's cluster, and allocates a fresh copy of the 221 * cluster with 1 ref and all the underlying chains locked. 222 * 223 * ip->cluster will be stable while the inode is locked. 224 * 225 * NOTE: We don't combine the inode/chain lock because putting away an 226 * inode would otherwise confuse multiple lock holders of the inode. 227 */ 228 void 229 hammer2_inode_lock(hammer2_inode_t *ip, int how) 230 { 231 hammer2_pfs_t *pmp; 232 233 hammer2_inode_ref(ip); 234 pmp = ip->pmp; 235 236 /* 237 * Inode structure mutex - Shared lock 238 */ 239 if (how & HAMMER2_RESOLVE_SHARED) { 240 hammer2_mtx_sh(&ip->lock); 241 return; 242 } 243 244 /* 245 * Inode structure mutex - Exclusive lock 246 * 247 * An exclusive lock (if not recursive) must wait for inodes on 248 * SYNCQ to flush first, to ensure that meta-data dependencies such 249 * as the nlink count and related directory entries are not split 250 * across flushes. 251 * 252 * If the vnode is locked by the current thread it must be unlocked 253 * across the tsleep() to avoid a deadlock. 254 */ 255 hammer2_mtx_ex(&ip->lock); 256 if (hammer2_mtx_refs(&ip->lock) > 1) 257 return; 258 while ((ip->flags & HAMMER2_INODE_SYNCQ) && pmp) { 259 hammer2_spin_ex(&pmp->list_spin); 260 if (ip->flags & HAMMER2_INODE_SYNCQ) { 261 tsleep_interlock(&ip->flags, 0); 262 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ_WAKEUP); 263 TAILQ_REMOVE(&pmp->syncq, ip, entry); 264 TAILQ_INSERT_HEAD(&pmp->syncq, ip, entry); 265 hammer2_spin_unex(&pmp->list_spin); 266 hammer2_mtx_unlock(&ip->lock); 267 tsleep(&ip->flags, PINTERLOCKED, "h2sync", 0); 268 hammer2_mtx_ex(&ip->lock); 269 continue; 270 } 271 hammer2_spin_unex(&pmp->list_spin); 272 break; 273 } 274 } 275 276 /* 277 * Exclusively lock up to four inodes, in order, with SYNCQ semantics. 278 * ip1 and ip2 must not be NULL. ip3 and ip4 may be NULL, but if ip3 is 279 * NULL then ip4 must also be NULL. 280 * 281 * This creates a dependency between up to four inodes. 282 */ 283 void 284 hammer2_inode_lock4(hammer2_inode_t *ip1, hammer2_inode_t *ip2, 285 hammer2_inode_t *ip3, hammer2_inode_t *ip4) 286 { 287 hammer2_inode_t *ips[4]; 288 hammer2_inode_t *iptmp; 289 hammer2_inode_t *ipslp; 290 hammer2_depend_t *depend; 291 hammer2_pfs_t *pmp; 292 size_t count; 293 size_t i; 294 295 pmp = ip1->pmp; /* may be NULL */ 296 KKASSERT(pmp == ip2->pmp); 297 298 ips[0] = ip1; 299 ips[1] = ip2; 300 if (ip3 == NULL) { 301 count = 2; 302 } else if (ip4 == NULL) { 303 count = 3; 304 ips[2] = ip3; 305 KKASSERT(pmp == ip3->pmp); 306 } else { 307 count = 4; 308 ips[2] = ip3; 309 ips[3] = ip4; 310 KKASSERT(pmp == ip3->pmp); 311 KKASSERT(pmp == ip4->pmp); 312 } 313 314 for (i = 0; i < count; ++i) 315 hammer2_inode_ref(ips[i]); 316 317 restart: 318 /* 319 * Lock the inodes in order 320 */ 321 for (i = 0; i < count; ++i) { 322 hammer2_mtx_ex(&ips[i]->lock); 323 } 324 325 /* 326 * Associate dependencies, record the first inode found on SYNCQ 327 * (operation is allowed to proceed for inodes on PASS2) for our 328 * sleep operation, this inode is theoretically the last one sync'd 329 * in the sequence. 330 * 331 * All inodes found on SYNCQ are moved to the head of the syncq 332 * to reduce stalls. 333 */ 334 hammer2_spin_ex(&pmp->list_spin); 335 depend = NULL; 336 ipslp = NULL; 337 for (i = 0; i < count; ++i) { 338 iptmp = ips[i]; 339 depend = hammer2_inode_setdepend_locked(iptmp, depend); 340 if (iptmp->flags & HAMMER2_INODE_SYNCQ) { 341 TAILQ_REMOVE(&pmp->syncq, iptmp, entry); 342 TAILQ_INSERT_HEAD(&pmp->syncq, iptmp, entry); 343 if (ipslp == NULL) 344 ipslp = iptmp; 345 } 346 } 347 hammer2_spin_unex(&pmp->list_spin); 348 349 /* 350 * Block and retry if any of the inodes are on SYNCQ. It is 351 * important that we allow the operation to proceed in the 352 * PASS2 case, to avoid deadlocking against the vnode. 353 */ 354 if (ipslp) { 355 for (i = 0; i < count; ++i) 356 hammer2_mtx_unlock(&ips[i]->lock); 357 tsleep(&ipslp->flags, 0, "h2sync", 2); 358 goto restart; 359 } 360 } 361 362 /* 363 * Release an inode lock. If another thread is blocked on SYNCQ_WAKEUP 364 * we wake them up. 365 */ 366 void 367 hammer2_inode_unlock(hammer2_inode_t *ip) 368 { 369 if (ip->flags & HAMMER2_INODE_SYNCQ_WAKEUP) { 370 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_WAKEUP); 371 hammer2_mtx_unlock(&ip->lock); 372 wakeup(&ip->flags); 373 } else { 374 hammer2_mtx_unlock(&ip->lock); 375 } 376 hammer2_inode_drop(ip); 377 } 378 379 /* 380 * If either ip1 or ip2 have been tapped by the syncer, make sure that both 381 * are. This ensure that dependencies (e.g. dirent-v-inode) are synced 382 * together. For dirent-v-inode depends, pass the dirent as ip1. 383 * 384 * If neither ip1 or ip2 have been tapped by the syncer, merge them into a 385 * single dependency. Dependencies are entered into pmp->depq. This 386 * effectively flags the inodes SIDEQ. 387 * 388 * Both ip1 and ip2 must be locked by the caller. This also ensures 389 * that we can't race the end of the syncer's queue run. 390 */ 391 void 392 hammer2_inode_depend(hammer2_inode_t *ip1, hammer2_inode_t *ip2) 393 { 394 hammer2_pfs_t *pmp; 395 hammer2_depend_t *depend; 396 397 pmp = ip1->pmp; 398 hammer2_spin_ex(&pmp->list_spin); 399 depend = hammer2_inode_setdepend_locked(ip1, NULL); 400 depend = hammer2_inode_setdepend_locked(ip2, depend); 401 hammer2_spin_unex(&pmp->list_spin); 402 } 403 404 /* 405 * Select a chain out of an inode's cluster and lock it. 406 * 407 * The inode does not have to be locked. 408 */ 409 hammer2_chain_t * 410 hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how) 411 { 412 hammer2_chain_t *chain; 413 hammer2_cluster_t *cluster; 414 415 hammer2_spin_sh(&ip->cluster_spin); 416 cluster = &ip->cluster; 417 if (clindex >= cluster->nchains) 418 chain = NULL; 419 else 420 chain = cluster->array[clindex].chain; 421 if (chain) { 422 hammer2_chain_ref(chain); 423 hammer2_spin_unsh(&ip->cluster_spin); 424 hammer2_chain_lock(chain, how); 425 } else { 426 hammer2_spin_unsh(&ip->cluster_spin); 427 } 428 return chain; 429 } 430 431 hammer2_chain_t * 432 hammer2_inode_chain_and_parent(hammer2_inode_t *ip, int clindex, 433 hammer2_chain_t **parentp, int how) 434 { 435 hammer2_chain_t *chain; 436 hammer2_chain_t *parent; 437 438 for (;;) { 439 hammer2_spin_sh(&ip->cluster_spin); 440 if (clindex >= ip->cluster.nchains) 441 chain = NULL; 442 else 443 chain = ip->cluster.array[clindex].chain; 444 if (chain) { 445 hammer2_chain_ref(chain); 446 hammer2_spin_unsh(&ip->cluster_spin); 447 hammer2_chain_lock(chain, how); 448 } else { 449 hammer2_spin_unsh(&ip->cluster_spin); 450 } 451 452 /* 453 * Get parent, lock order must be (parent, chain). 454 */ 455 parent = chain->parent; 456 if (parent) { 457 hammer2_chain_ref(parent); 458 hammer2_chain_unlock(chain); 459 hammer2_chain_lock(parent, how); 460 hammer2_chain_lock(chain, how); 461 } 462 if (ip->cluster.array[clindex].chain == chain && 463 chain->parent == parent) { 464 break; 465 } 466 467 /* 468 * Retry 469 */ 470 hammer2_chain_unlock(chain); 471 hammer2_chain_drop(chain); 472 if (parent) { 473 hammer2_chain_unlock(parent); 474 hammer2_chain_drop(parent); 475 } 476 } 477 *parentp = parent; 478 479 return chain; 480 } 481 482 /* 483 * Temporarily release a lock held shared or exclusive. Caller must 484 * hold the lock shared or exclusive on call and lock will be released 485 * on return. 486 * 487 * Restore a lock that was temporarily released. 488 */ 489 hammer2_mtx_state_t 490 hammer2_inode_lock_temp_release(hammer2_inode_t *ip) 491 { 492 return hammer2_mtx_temp_release(&ip->lock); 493 } 494 495 void 496 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, hammer2_mtx_state_t ostate) 497 { 498 hammer2_mtx_temp_restore(&ip->lock, ostate); 499 } 500 501 /* 502 * Upgrade a shared inode lock to exclusive and return. If the inode lock 503 * is already held exclusively this is a NOP. 504 * 505 * The caller MUST hold the inode lock either shared or exclusive on call 506 * and will own the lock exclusively on return. 507 * 508 * Returns non-zero if the lock was already exclusive prior to the upgrade. 509 */ 510 int 511 hammer2_inode_lock_upgrade(hammer2_inode_t *ip) 512 { 513 int wasexclusive; 514 515 /* XXX pretends it wasn't exclusive, but shouldn't matter */ 516 //if (mtx_islocked_ex(&ip->lock)) { 517 if (0) { 518 wasexclusive = 1; 519 } else { 520 hammer2_mtx_unlock(&ip->lock); 521 hammer2_mtx_ex(&ip->lock); 522 wasexclusive = 0; 523 } 524 return wasexclusive; 525 } 526 527 /* 528 * Downgrade an inode lock from exclusive to shared only if the inode 529 * lock was previously shared. If the inode lock was previously exclusive, 530 * this is a NOP. 531 */ 532 void 533 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int wasexclusive) 534 { 535 if (wasexclusive == 0) 536 hammer2_mtx_downgrade(&ip->lock); 537 } 538 539 /* 540 * Lookup an inode by inode number 541 */ 542 hammer2_inode_t * 543 hammer2_inode_lookup(hammer2_pfs_t *pmp, hammer2_tid_t inum) 544 { 545 hammer2_inode_t *ip; 546 547 KKASSERT(pmp); 548 if (pmp->spmp_hmp) { 549 ip = NULL; 550 } else { 551 hammer2_spin_ex(&pmp->inum_spin); 552 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum); 553 if (ip) 554 hammer2_inode_ref(ip); 555 hammer2_spin_unex(&pmp->inum_spin); 556 } 557 return(ip); 558 } 559 560 /* 561 * Adding a ref to an inode is only legal if the inode already has at least 562 * one ref. 563 * 564 * (can be called with spinlock held) 565 */ 566 void 567 hammer2_inode_ref(hammer2_inode_t *ip) 568 { 569 atomic_add_int(&ip->refs, 1); 570 if (hammer2_debug & 0x80000) { 571 kprintf("INODE+1 %p (%d->%d)\n", ip, ip->refs - 1, ip->refs); 572 print_backtrace(8); 573 } 574 } 575 576 /* 577 * Drop an inode reference, freeing the inode when the last reference goes 578 * away. 579 */ 580 void 581 hammer2_inode_drop(hammer2_inode_t *ip) 582 { 583 hammer2_pfs_t *pmp; 584 u_int refs; 585 586 while (ip) { 587 if (hammer2_debug & 0x80000) { 588 kprintf("INODE-1 %p (%d->%d)\n", 589 ip, ip->refs, ip->refs - 1); 590 print_backtrace(8); 591 } 592 refs = ip->refs; 593 cpu_ccfence(); 594 if (refs == 1) { 595 /* 596 * Transition to zero, must interlock with 597 * the inode inumber lookup tree (if applicable). 598 * It should not be possible for anyone to race 599 * the transition to 0. 600 */ 601 pmp = ip->pmp; 602 KKASSERT(pmp); 603 hammer2_spin_ex(&pmp->inum_spin); 604 605 if (atomic_cmpset_int(&ip->refs, 1, 0)) { 606 KKASSERT(hammer2_mtx_refs(&ip->lock) == 0); 607 if (ip->flags & HAMMER2_INODE_ONRBTREE) { 608 atomic_clear_int(&ip->flags, 609 HAMMER2_INODE_ONRBTREE); 610 RB_REMOVE(hammer2_inode_tree, 611 &pmp->inum_tree, ip); 612 --pmp->inum_count; 613 } 614 hammer2_spin_unex(&pmp->inum_spin); 615 616 ip->pmp = NULL; 617 618 /* 619 * Cleaning out ip->cluster isn't entirely 620 * trivial. 621 */ 622 hammer2_inode_repoint(ip, NULL); 623 624 /* 625 * VOP_RECLAIM is currently unused, 626 * so directly free vnode before inode. 627 */ 628 if (ip->vp) { 629 if (ip->vp->v_malloced) 630 freevnode(ip->vp); 631 } else { 632 /* PFS inode ? */ 633 } 634 635 kfree_obj(ip, pmp->minode); 636 atomic_add_long(&pmp->inmem_inodes, -1); 637 ip = NULL; /* will terminate loop */ 638 } else { 639 hammer2_spin_unex(&ip->pmp->inum_spin); 640 } 641 } else { 642 /* 643 * Non zero transition 644 */ 645 if (atomic_cmpset_int(&ip->refs, refs, refs - 1)) 646 break; 647 } 648 } 649 } 650 651 /* 652 * Get the vnode associated with the given inode, allocating the vnode if 653 * necessary. The vnode will be returned exclusively locked. 654 * 655 * *errorp is set to a UNIX error, not a HAMMER2 error. 656 * 657 * The caller must lock the inode (shared or exclusive). 658 * 659 * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim 660 * races. 661 */ 662 struct m_vnode * 663 hammer2_igetv(hammer2_inode_t *ip, int *errorp) 664 { 665 hammer2_pfs_t *pmp; 666 struct m_vnode *vp; 667 668 pmp = ip->pmp; 669 KKASSERT(pmp != NULL); 670 *errorp = 0; 671 672 for (;;) { 673 /* 674 * Attempt to reuse an existing vnode assignment. It is 675 * possible to race a reclaim so the vget() may fail. The 676 * inode must be unlocked during the vget() to avoid a 677 * deadlock against a reclaim. 678 */ 679 int wasexclusive; 680 681 vp = ip->vp; 682 if (vp) { 683 /* 684 * Inode must be unlocked during the vget() to avoid 685 * possible deadlocks, but leave the ip ref intact. 686 * 687 * vnode is held to prevent destruction during the 688 * vget(). The vget() can still fail if we lost 689 * a reclaim race on the vnode. 690 */ 691 hammer2_mtx_state_t ostate; 692 693 vhold(vp); 694 ostate = hammer2_inode_lock_temp_release(ip); 695 if (vget(vp, LK_EXCLUSIVE)) { 696 vdrop(vp); 697 hammer2_inode_lock_temp_restore(ip, ostate); 698 continue; 699 } 700 hammer2_inode_lock_temp_restore(ip, ostate); 701 vdrop(vp); 702 /* vp still locked and ref from vget */ 703 if (ip->vp != vp) { 704 kprintf("hammer2: igetv race %p/%p\n", 705 ip->vp, vp); 706 vput(vp); 707 continue; 708 } 709 *errorp = 0; 710 break; 711 } 712 713 /* 714 * No vnode exists, allocate a new vnode. Beware of 715 * allocation races. This function will return an 716 * exclusively locked and referenced vnode. 717 */ 718 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0); 719 if (*errorp) { 720 kprintf("hammer2: igetv getnewvnode failed %d\n", 721 *errorp); 722 vp = NULL; 723 break; 724 } 725 726 /* 727 * Lock the inode and check for an allocation race. 728 */ 729 wasexclusive = hammer2_inode_lock_upgrade(ip); 730 if (ip->vp != NULL) { 731 vp->v_type = VBAD; 732 vx_put(vp); 733 hammer2_inode_lock_downgrade(ip, wasexclusive); 734 continue; 735 } 736 737 switch (ip->meta.type) { 738 case HAMMER2_OBJTYPE_DIRECTORY: 739 vp->v_type = VDIR; 740 break; 741 case HAMMER2_OBJTYPE_REGFILE: 742 /* 743 * Regular file must use buffer cache I/O 744 * (VKVABIO cpu sync semantics supported) 745 */ 746 vp->v_type = VREG; 747 vsetflags(vp, VKVABIO); 748 vinitvmio(vp, ip->meta.size, 749 HAMMER2_LBUFSIZE, 750 (int)ip->meta.size & HAMMER2_LBUFMASK); 751 break; 752 case HAMMER2_OBJTYPE_SOFTLINK: 753 /* 754 * XXX for now we are using the generic file_read 755 * and file_write code so we need a buffer cache 756 * association. 757 * 758 * (VKVABIO cpu sync semantics supported) 759 */ 760 vp->v_type = VLNK; 761 vsetflags(vp, VKVABIO); 762 vinitvmio(vp, ip->meta.size, 763 HAMMER2_LBUFSIZE, 764 (int)ip->meta.size & HAMMER2_LBUFMASK); 765 break; 766 case HAMMER2_OBJTYPE_CDEV: 767 vp->v_type = VCHR; 768 /* fall through */ 769 case HAMMER2_OBJTYPE_BDEV: 770 //vp->v_ops = &pmp->mp->mnt_vn_spec_ops; 771 if (ip->meta.type != HAMMER2_OBJTYPE_CDEV) 772 vp->v_type = VBLK; 773 addaliasu(vp, 774 ip->meta.rmajor, 775 ip->meta.rminor); 776 break; 777 case HAMMER2_OBJTYPE_FIFO: 778 vp->v_type = VFIFO; 779 //vp->v_ops = &pmp->mp->mnt_vn_fifo_ops; 780 break; 781 case HAMMER2_OBJTYPE_SOCKET: 782 vp->v_type = VSOCK; 783 break; 784 default: 785 panic("hammer2: unhandled objtype %d", 786 ip->meta.type); 787 break; 788 } 789 790 if (ip == pmp->iroot) 791 vsetflags(vp, VROOT); 792 793 vp->v_data = ip; 794 ip->vp = vp; 795 hammer2_inode_ref(ip); /* vp association */ 796 hammer2_inode_lock_downgrade(ip, wasexclusive); 797 vx_downgrade(vp); 798 break; 799 } 800 801 /* 802 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0. 803 */ 804 if (hammer2_debug & 0x0002) { 805 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n", 806 vp, -1, -1); 807 } 808 return (vp); 809 } 810 811 /* 812 * XXX this API needs a rewrite. It needs to be split into a 813 * hammer2_inode_alloc() and hammer2_inode_build() to allow us to get 814 * rid of the inode/chain lock reversal fudge. 815 * 816 * Returns the inode associated with the passed-in cluster, allocating a new 817 * hammer2_inode structure if necessary, then synchronizing it to the passed 818 * xop cluster. When synchronizing, if idx >= 0, only cluster index (idx) 819 * is synchronized. Otherwise the whole cluster is synchronized. inum will 820 * be extracted from the passed-in xop and the inum argument will be ignored. 821 * 822 * If xop is passed as NULL then a new hammer2_inode is allocated with the 823 * specified inum, and returned. For normal inodes, the inode will be 824 * indexed in memory and if it already exists the existing ip will be 825 * returned instead of allocating a new one. The superroot and PFS inodes 826 * are not indexed in memory. 827 * 828 * The passed-in cluster must be locked and will remain locked on return. 829 * The returned inode will be locked and the caller may dispose of both 830 * via hammer2_inode_unlock() + hammer2_inode_drop(). However, if the caller 831 * needs to resolve a hardlink it must ref/unlock/relock/drop the inode. 832 * 833 * The hammer2_inode structure regulates the interface between the high level 834 * kernel VNOPS API and the filesystem backend (the chains). 835 * 836 * On return the inode is locked with the supplied cluster. 837 */ 838 hammer2_inode_t * 839 hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_xop_head_t *xop, 840 hammer2_tid_t inum, int idx) 841 { 842 hammer2_inode_t *nip; 843 const hammer2_inode_data_t *iptmp; 844 const hammer2_inode_data_t *nipdata; 845 846 KKASSERT(xop == NULL || 847 hammer2_cluster_type(&xop->cluster) == 848 HAMMER2_BREF_TYPE_INODE); 849 KKASSERT(pmp); 850 851 /* 852 * Interlocked lookup/ref of the inode. This code is only needed 853 * when looking up inodes with nlinks != 0 (TODO: optimize out 854 * otherwise and test for duplicates). 855 * 856 * Cluster can be NULL during the initial pfs allocation. 857 */ 858 if (xop) { 859 iptmp = &hammer2_xop_gdata(xop)->ipdata; 860 inum = iptmp->meta.inum; 861 hammer2_xop_pdata(xop); 862 } 863 again: 864 nip = hammer2_inode_lookup(pmp, inum); 865 if (nip) { 866 /* 867 * We may have to unhold the cluster to avoid a deadlock 868 * against vnlru (and possibly other XOPs). 869 */ 870 if (xop) { 871 if (hammer2_mtx_ex_try(&nip->lock) != 0) { 872 hammer2_cluster_unhold(&xop->cluster); 873 hammer2_mtx_ex(&nip->lock); 874 hammer2_cluster_rehold(&xop->cluster); 875 } 876 } else { 877 hammer2_mtx_ex(&nip->lock); 878 } 879 880 /* 881 * Handle SMP race (not applicable to the super-root spmp 882 * which can't index inodes due to duplicative inode numbers). 883 */ 884 if (pmp->spmp_hmp == NULL && 885 (nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { 886 hammer2_mtx_unlock(&nip->lock); 887 hammer2_inode_drop(nip); 888 goto again; 889 } 890 if (xop) { 891 if (idx >= 0) 892 hammer2_inode_repoint_one(nip, &xop->cluster, 893 idx); 894 else 895 hammer2_inode_repoint(nip, &xop->cluster); 896 } 897 return nip; 898 } 899 900 /* 901 * We couldn't find the inode number, create a new inode and try to 902 * insert it, handle insertion races. 903 */ 904 nip = kmalloc_obj(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO); 905 hammer2_spin_init(&nip->cluster_spin, "h2clspin"); 906 atomic_add_long(&pmp->inmem_inodes, 1); 907 908 /* 909 * Initialize nip's cluster. A cluster is provided for normal 910 * inodes but typically not for the super-root or PFS inodes. 911 */ 912 { 913 hammer2_inode_t *nnip = nip; 914 nip->ihash = (int)hammer2_icrc32(&nnip, sizeof(nnip)); 915 } 916 917 nip->cluster.refs = 1; 918 nip->cluster.pmp = pmp; 919 nip->cluster.flags |= HAMMER2_CLUSTER_INODE; 920 if (xop) { 921 nipdata = &hammer2_xop_gdata(xop)->ipdata; 922 nip->meta = nipdata->meta; 923 hammer2_xop_pdata(xop); 924 hammer2_inode_repoint(nip, &xop->cluster); 925 } else { 926 nip->meta.inum = inum; /* PFS inum is always 1 XXX */ 927 /* mtime will be updated when a cluster is available */ 928 } 929 930 nip->pmp = pmp; 931 932 /* 933 * ref and lock on nip gives it state compatible to after a 934 * hammer2_inode_lock() call. 935 */ 936 nip->refs = 1; 937 hammer2_mtx_init(&nip->lock, "h2inode"); 938 hammer2_mtx_init(&nip->truncate_lock, "h2trunc"); 939 hammer2_mtx_ex(&nip->lock); 940 TAILQ_INIT(&nip->depend_static.sideq); 941 /* combination of thread lock and chain lock == inode lock */ 942 943 /* 944 * Attempt to add the inode. If it fails we raced another inode 945 * get. Undo all the work and try again. 946 */ 947 if (pmp->spmp_hmp == NULL) { 948 hammer2_spin_ex(&pmp->inum_spin); 949 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) { 950 hammer2_spin_unex(&pmp->inum_spin); 951 hammer2_mtx_unlock(&nip->lock); 952 hammer2_inode_drop(nip); 953 goto again; 954 } 955 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE); 956 ++pmp->inum_count; 957 hammer2_spin_unex(&pmp->inum_spin); 958 } 959 return (nip); 960 } 961 962 /* 963 * Create a PFS inode under the superroot. This function will create the 964 * inode, its media chains, and also insert it into the media. 965 * 966 * Caller must be in a flush transaction because we are inserting the inode 967 * onto the media. 968 */ 969 hammer2_inode_t * 970 hammer2_inode_create_pfs(hammer2_pfs_t *spmp, 971 const char *name, size_t name_len, 972 int *errorp) 973 { 974 hammer2_xop_create_t *xop; 975 hammer2_inode_t *pip; 976 hammer2_inode_t *nip; 977 int error; 978 uint8_t pip_comp_algo; 979 uint8_t pip_check_algo; 980 hammer2_tid_t pip_inum; 981 hammer2_key_t lhc; 982 983 pip = spmp->iroot; 984 nip = NULL; 985 986 lhc = hammer2_dirhash(name, name_len); 987 *errorp = 0; 988 989 /* 990 * Locate the inode or indirect block to create the new 991 * entry in. At the same time check for key collisions 992 * and iterate until we don't get one. 993 * 994 * Lock the directory exclusively for now to guarantee that 995 * we can find an unused lhc for the name. Due to collisions, 996 * two different creates can end up with the same lhc so we 997 * cannot depend on the OS to prevent the collision. 998 */ 999 hammer2_inode_lock(pip, 0); 1000 1001 pip_comp_algo = pip->meta.comp_algo; 1002 pip_check_algo = pip->meta.check_algo; 1003 pip_inum = (pip == pip->pmp->iroot) ? 1 : pip->meta.inum; 1004 1005 /* 1006 * Locate an unused key in the collision space. 1007 */ 1008 { 1009 hammer2_xop_scanlhc_t *sxop; 1010 hammer2_key_t lhcbase; 1011 1012 lhcbase = lhc; 1013 sxop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 1014 sxop->lhc = lhc; 1015 hammer2_xop_start(&sxop->head, &hammer2_scanlhc_desc); 1016 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) { 1017 if (lhc != sxop->head.cluster.focus->bref.key) 1018 break; 1019 ++lhc; 1020 } 1021 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP); 1022 1023 if (error) { 1024 if (error != HAMMER2_ERROR_ENOENT) 1025 goto done2; 1026 ++lhc; 1027 error = 0; 1028 } 1029 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) { 1030 error = HAMMER2_ERROR_ENOSPC; 1031 goto done2; 1032 } 1033 } 1034 1035 /* 1036 * Create the inode with the lhc as the key. 1037 */ 1038 xop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 1039 xop->lhc = lhc; 1040 xop->flags = HAMMER2_INSERT_PFSROOT; 1041 bzero(&xop->meta, sizeof(xop->meta)); 1042 1043 xop->meta.type = HAMMER2_OBJTYPE_DIRECTORY; 1044 xop->meta.inum = 1; 1045 xop->meta.iparent = pip_inum; 1046 1047 /* Inherit parent's inode compression mode. */ 1048 xop->meta.comp_algo = pip_comp_algo; 1049 xop->meta.check_algo = pip_check_algo; 1050 xop->meta.version = HAMMER2_INODE_VERSION_ONE; 1051 hammer2_update_time(&xop->meta.ctime); 1052 xop->meta.mtime = xop->meta.ctime; 1053 xop->meta.mode = 0755; 1054 xop->meta.nlinks = 1; 1055 1056 /* 1057 * Regular files and softlinks allow a small amount of data to be 1058 * directly embedded in the inode. This flag will be cleared if 1059 * the size is extended past the embedded limit. 1060 */ 1061 if (xop->meta.type == HAMMER2_OBJTYPE_REGFILE || 1062 xop->meta.type == HAMMER2_OBJTYPE_SOFTLINK) { 1063 xop->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA; 1064 } 1065 hammer2_xop_setname(&xop->head, name, name_len); 1066 xop->meta.name_len = name_len; 1067 xop->meta.name_key = lhc; 1068 KKASSERT(name_len < HAMMER2_INODE_MAXNAME); 1069 1070 hammer2_xop_start(&xop->head, &hammer2_inode_create_desc); 1071 1072 error = hammer2_xop_collect(&xop->head, 0); 1073 #if INODE_DEBUG 1074 kprintf("CREATE INODE %*.*s\n", 1075 (int)name_len, (int)name_len, name); 1076 #endif 1077 1078 if (error) { 1079 *errorp = error; 1080 goto done; 1081 } 1082 1083 /* 1084 * Set up the new inode if not a hardlink pointer. 1085 * 1086 * NOTE: *_get() integrates chain's lock into the inode lock. 1087 * 1088 * NOTE: Only one new inode can currently be created per 1089 * transaction. If the need arises we can adjust 1090 * hammer2_trans_init() to allow more. 1091 * 1092 * NOTE: nipdata will have chain's blockset data. 1093 */ 1094 nip = hammer2_inode_get(pip->pmp, &xop->head, -1, -1); 1095 nip->comp_heuristic = 0; 1096 done: 1097 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1098 done2: 1099 hammer2_inode_unlock(pip); 1100 1101 return (nip); 1102 } 1103 1104 /* 1105 * Create a new, normal inode. This function will create the inode, 1106 * the media chains, but will not insert the chains onto the media topology 1107 * (doing so would require a flush transaction and cause long stalls). 1108 * 1109 * Caller must be in a normal transaction. 1110 */ 1111 hammer2_inode_t * 1112 hammer2_inode_create_normal(hammer2_inode_t *pip, 1113 struct vattr *vap, struct ucred *cred, 1114 hammer2_key_t inum, int *errorp) 1115 { 1116 hammer2_xop_create_t *xop; 1117 hammer2_inode_t *dip; 1118 hammer2_inode_t *nip; 1119 int error; 1120 uid_t xuid; 1121 uuid_t pip_uid; 1122 uuid_t pip_gid; 1123 uint32_t pip_mode; 1124 uint8_t pip_comp_algo; 1125 uint8_t pip_check_algo; 1126 hammer2_tid_t pip_inum; 1127 1128 dip = pip->pmp->iroot; 1129 KKASSERT(dip != NULL); 1130 1131 *errorp = 0; 1132 1133 /*hammer2_inode_lock(dip, 0);*/ 1134 1135 pip_uid = pip->meta.uid; 1136 pip_gid = pip->meta.gid; 1137 pip_mode = pip->meta.mode; 1138 pip_comp_algo = pip->meta.comp_algo; 1139 pip_check_algo = pip->meta.check_algo; 1140 pip_inum = (pip == pip->pmp->iroot) ? 1 : pip->meta.inum; 1141 1142 /* 1143 * Create the in-memory hammer2_inode structure for the specified 1144 * inode. 1145 */ 1146 nip = hammer2_inode_get(dip->pmp, NULL, inum, -1); 1147 nip->comp_heuristic = 0; 1148 KKASSERT((nip->flags & HAMMER2_INODE_CREATING) == 0 && 1149 nip->cluster.nchains == 0); 1150 atomic_set_int(&nip->flags, HAMMER2_INODE_CREATING); 1151 1152 /* 1153 * Setup the inode meta-data 1154 */ 1155 nip->meta.type = hammer2_get_obj_type(vap->va_type); 1156 1157 switch (nip->meta.type) { 1158 case HAMMER2_OBJTYPE_CDEV: 1159 case HAMMER2_OBJTYPE_BDEV: 1160 assert(0); /* XXX unsupported */ 1161 nip->meta.rmajor = vap->va_rmajor; 1162 nip->meta.rminor = vap->va_rminor; 1163 break; 1164 default: 1165 break; 1166 } 1167 1168 KKASSERT(nip->meta.inum == inum); 1169 nip->meta.iparent = pip_inum; 1170 1171 /* Inherit parent's inode compression mode. */ 1172 nip->meta.comp_algo = pip_comp_algo; 1173 nip->meta.check_algo = pip_check_algo; 1174 nip->meta.version = HAMMER2_INODE_VERSION_ONE; 1175 hammer2_update_time(&nip->meta.ctime); 1176 nip->meta.mtime = nip->meta.ctime; 1177 nip->meta.mode = vap->va_mode; 1178 nip->meta.nlinks = 1; 1179 1180 xuid = hammer2_to_unix_xid(&pip_uid); 1181 xuid = vop_helper_create_uid(dip->pmp->mp, pip_mode, 1182 xuid, cred, 1183 &vap->va_mode); 1184 if (vap->va_vaflags & VA_UID_UUID_VALID) 1185 nip->meta.uid = vap->va_uid_uuid; 1186 else if (vap->va_uid != (uid_t)VNOVAL) 1187 hammer2_guid_to_uuid(&nip->meta.uid, vap->va_uid); 1188 else 1189 hammer2_guid_to_uuid(&nip->meta.uid, xuid); 1190 1191 if (vap->va_vaflags & VA_GID_UUID_VALID) 1192 nip->meta.gid = vap->va_gid_uuid; 1193 else if (vap->va_gid != (gid_t)VNOVAL) 1194 hammer2_guid_to_uuid(&nip->meta.gid, vap->va_gid); 1195 else 1196 nip->meta.gid = pip_gid; 1197 1198 /* 1199 * Regular files and softlinks allow a small amount of data to be 1200 * directly embedded in the inode. This flag will be cleared if 1201 * the size is extended past the embedded limit. 1202 */ 1203 if (nip->meta.type == HAMMER2_OBJTYPE_REGFILE || 1204 nip->meta.type == HAMMER2_OBJTYPE_SOFTLINK) { 1205 nip->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA; 1206 } 1207 1208 /* 1209 * Create the inode using (inum) as the key. Pass pip for 1210 * method inheritance. 1211 */ 1212 xop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 1213 xop->lhc = inum; 1214 xop->flags = 0; 1215 xop->meta = nip->meta; 1216 KKASSERT(vap); 1217 1218 xop->meta.name_len = hammer2_xop_setname_inum(&xop->head, inum); 1219 xop->meta.name_key = inum; 1220 nip->meta.name_len = xop->meta.name_len; 1221 nip->meta.name_key = xop->meta.name_key; 1222 hammer2_inode_modify(nip); 1223 1224 /* 1225 * Create the inode media chains but leave them detached. We are 1226 * not in a flush transaction so we can't mess with media topology 1227 * above normal inodes (i.e. the index of the inodes themselves). 1228 * 1229 * We've already set the INODE_CREATING flag. The inode's media 1230 * chains will be inserted onto the media topology on the next 1231 * filesystem sync. 1232 */ 1233 hammer2_xop_start(&xop->head, &hammer2_inode_create_det_desc); 1234 1235 error = hammer2_xop_collect(&xop->head, 0); 1236 #if INODE_DEBUG 1237 kprintf("create inode type %d error %d\n", nip->meta.type, error); 1238 #endif 1239 1240 if (error) { 1241 *errorp = error; 1242 goto done; 1243 } 1244 1245 /* 1246 * Associate the media chains created by the backend with the 1247 * frontend inode. 1248 */ 1249 hammer2_inode_repoint(nip, &xop->head.cluster); 1250 done: 1251 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1252 /*hammer2_inode_unlock(dip);*/ 1253 1254 return (nip); 1255 } 1256 1257 /* 1258 * Create a directory entry under dip with the specified name, inode number, 1259 * and OBJTYPE (type). 1260 * 1261 * This returns a UNIX errno code, not a HAMMER2_ERROR_* code. 1262 * 1263 * Caller must hold dip locked. 1264 */ 1265 int 1266 hammer2_dirent_create(hammer2_inode_t *dip, const char *name, size_t name_len, 1267 hammer2_key_t inum, uint8_t type) 1268 { 1269 hammer2_xop_mkdirent_t *xop; 1270 hammer2_key_t lhc; 1271 int error; 1272 1273 lhc = 0; 1274 error = 0; 1275 1276 KKASSERT(name != NULL); 1277 lhc = hammer2_dirhash(name, name_len); 1278 1279 /* 1280 * Locate the inode or indirect block to create the new 1281 * entry in. At the same time check for key collisions 1282 * and iterate until we don't get one. 1283 * 1284 * Lock the directory exclusively for now to guarantee that 1285 * we can find an unused lhc for the name. Due to collisions, 1286 * two different creates can end up with the same lhc so we 1287 * cannot depend on the OS to prevent the collision. 1288 */ 1289 hammer2_inode_modify(dip); 1290 1291 /* 1292 * If name specified, locate an unused key in the collision space. 1293 * Otherwise use the passed-in lhc directly. 1294 */ 1295 { 1296 hammer2_xop_scanlhc_t *sxop; 1297 hammer2_key_t lhcbase; 1298 1299 lhcbase = lhc; 1300 sxop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING); 1301 sxop->lhc = lhc; 1302 hammer2_xop_start(&sxop->head, &hammer2_scanlhc_desc); 1303 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) { 1304 if (lhc != sxop->head.cluster.focus->bref.key) 1305 break; 1306 ++lhc; 1307 } 1308 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP); 1309 1310 if (error) { 1311 if (error != HAMMER2_ERROR_ENOENT) 1312 goto done2; 1313 ++lhc; 1314 error = 0; 1315 } 1316 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) { 1317 error = HAMMER2_ERROR_ENOSPC; 1318 goto done2; 1319 } 1320 } 1321 1322 /* 1323 * Create the directory entry with the lhc as the key. 1324 */ 1325 xop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING); 1326 xop->lhc = lhc; 1327 bzero(&xop->dirent, sizeof(xop->dirent)); 1328 xop->dirent.inum = inum; 1329 xop->dirent.type = type; 1330 xop->dirent.namlen = name_len; 1331 1332 KKASSERT(name_len < HAMMER2_INODE_MAXNAME); 1333 hammer2_xop_setname(&xop->head, name, name_len); 1334 1335 hammer2_xop_start(&xop->head, &hammer2_inode_mkdirent_desc); 1336 1337 error = hammer2_xop_collect(&xop->head, 0); 1338 1339 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1340 done2: 1341 error = hammer2_error_to_errno(error); 1342 1343 return error; 1344 } 1345 1346 /* 1347 * Repoint ip->cluster's chains to cluster's chains and fixup the default 1348 * focus. All items, valid or invalid, are repointed. hammer2_xop_start() 1349 * filters out invalid or non-matching elements. 1350 * 1351 * Caller must hold the inode and cluster exclusive locked, if not NULL, 1352 * must also be locked. 1353 * 1354 * Cluster may be NULL to clean out any chains in ip->cluster. 1355 */ 1356 void 1357 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_cluster_t *cluster) 1358 { 1359 hammer2_chain_t *dropch[HAMMER2_MAXCLUSTER]; 1360 hammer2_chain_t *ochain; 1361 hammer2_chain_t *nchain; 1362 int i; 1363 1364 bzero(dropch, sizeof(dropch)); 1365 1366 /* 1367 * Replace chains in ip->cluster with chains from cluster and 1368 * adjust the focus if necessary. 1369 * 1370 * NOTE: nchain and/or ochain can be NULL due to gaps 1371 * in the cluster arrays. 1372 */ 1373 hammer2_spin_ex(&ip->cluster_spin); 1374 for (i = 0; cluster && i < cluster->nchains; ++i) { 1375 /* 1376 * Do not replace elements which are the same. Also handle 1377 * element count discrepancies. 1378 */ 1379 nchain = cluster->array[i].chain; 1380 if (i < ip->cluster.nchains) { 1381 ochain = ip->cluster.array[i].chain; 1382 if (ochain == nchain) 1383 continue; 1384 } else { 1385 ochain = NULL; 1386 } 1387 1388 /* 1389 * Make adjustments 1390 */ 1391 ip->cluster.array[i].chain = nchain; 1392 ip->cluster.array[i].flags &= ~HAMMER2_CITEM_INVALID; 1393 ip->cluster.array[i].flags |= cluster->array[i].flags & 1394 HAMMER2_CITEM_INVALID; 1395 if (nchain) 1396 hammer2_chain_ref(nchain); 1397 dropch[i] = ochain; 1398 } 1399 1400 /* 1401 * Release any left-over chains in ip->cluster. 1402 */ 1403 while (i < ip->cluster.nchains) { 1404 nchain = ip->cluster.array[i].chain; 1405 if (nchain) { 1406 ip->cluster.array[i].chain = NULL; 1407 ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID; 1408 } 1409 dropch[i] = nchain; 1410 ++i; 1411 } 1412 1413 /* 1414 * Fixup fields. Note that the inode-embedded cluster is never 1415 * directly locked. 1416 */ 1417 if (cluster) { 1418 ip->cluster.nchains = cluster->nchains; 1419 ip->cluster.focus = cluster->focus; 1420 ip->cluster.flags = cluster->flags & ~HAMMER2_CLUSTER_LOCKED; 1421 } else { 1422 ip->cluster.nchains = 0; 1423 ip->cluster.focus = NULL; 1424 ip->cluster.flags &= ~HAMMER2_CLUSTER_ZFLAGS; 1425 } 1426 1427 hammer2_spin_unex(&ip->cluster_spin); 1428 1429 /* 1430 * Cleanup outside of spinlock 1431 */ 1432 while (--i >= 0) { 1433 if (dropch[i]) 1434 hammer2_chain_drop(dropch[i]); 1435 } 1436 } 1437 1438 /* 1439 * Repoint a single element from the cluster to the ip. Used by the 1440 * synchronization threads to piecemeal update inodes. Does not change 1441 * focus and requires inode to be re-locked to clean-up flags (XXX). 1442 */ 1443 void 1444 hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster, 1445 int idx) 1446 { 1447 hammer2_chain_t *ochain; 1448 hammer2_chain_t *nchain; 1449 int i; 1450 1451 hammer2_spin_ex(&ip->cluster_spin); 1452 KKASSERT(idx < cluster->nchains); 1453 if (idx < ip->cluster.nchains) { 1454 ochain = ip->cluster.array[idx].chain; 1455 nchain = cluster->array[idx].chain; 1456 } else { 1457 ochain = NULL; 1458 nchain = cluster->array[idx].chain; 1459 for (i = ip->cluster.nchains; i <= idx; ++i) { 1460 bzero(&ip->cluster.array[i], 1461 sizeof(ip->cluster.array[i])); 1462 ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID; 1463 } 1464 ip->cluster.nchains = idx + 1; 1465 } 1466 if (ochain != nchain) { 1467 /* 1468 * Make adjustments. 1469 */ 1470 ip->cluster.array[idx].chain = nchain; 1471 ip->cluster.array[idx].flags &= ~HAMMER2_CITEM_INVALID; 1472 ip->cluster.array[idx].flags |= cluster->array[idx].flags & 1473 HAMMER2_CITEM_INVALID; 1474 } 1475 hammer2_spin_unex(&ip->cluster_spin); 1476 if (ochain != nchain) { 1477 if (nchain) 1478 hammer2_chain_ref(nchain); 1479 if (ochain) 1480 hammer2_chain_drop(ochain); 1481 } 1482 } 1483 1484 hammer2_key_t 1485 hammer2_inode_data_count(const hammer2_inode_t *ip) 1486 { 1487 hammer2_chain_t *chain; 1488 hammer2_key_t count = 0; 1489 int i; 1490 1491 for (i = 0; i < ip->cluster.nchains; ++i) { 1492 if ((chain = ip->cluster.array[i].chain) != NULL) { 1493 if (count < chain->bref.embed.stats.data_count) 1494 count = chain->bref.embed.stats.data_count; 1495 } 1496 } 1497 return count; 1498 } 1499 1500 hammer2_key_t 1501 hammer2_inode_inode_count(const hammer2_inode_t *ip) 1502 { 1503 hammer2_chain_t *chain; 1504 hammer2_key_t count = 0; 1505 int i; 1506 1507 for (i = 0; i < ip->cluster.nchains; ++i) { 1508 if ((chain = ip->cluster.array[i].chain) != NULL) { 1509 if (count < chain->bref.embed.stats.inode_count) 1510 count = chain->bref.embed.stats.inode_count; 1511 } 1512 } 1513 return count; 1514 } 1515 1516 /* 1517 * Called with a locked inode to finish unlinking an inode after xop_unlink 1518 * had been run. This function is responsible for decrementing nlinks. 1519 * 1520 * We don't bother decrementing nlinks if the file is not open and this was 1521 * the last link. 1522 * 1523 * If the inode is a hardlink target it's chain has not yet been deleted, 1524 * otherwise it's chain has been deleted. 1525 * 1526 * If isopen then any prior deletion was not permanent and the inode is 1527 * left intact with nlinks == 0; 1528 */ 1529 int 1530 hammer2_inode_unlink_finisher(hammer2_inode_t *ip, struct m_vnode **vprecyclep) 1531 { 1532 struct m_vnode *vp; 1533 1534 /* 1535 * Decrement nlinks. Catch a bad nlinks count here too (e.g. 0 or 1536 * negative), and just assume a transition to 0. 1537 */ 1538 if ((int64_t)ip->meta.nlinks <= 1) { 1539 atomic_set_int(&ip->flags, HAMMER2_INODE_ISUNLINKED); 1540 1541 /* 1542 * Scrap the vnode as quickly as possible. The vp association 1543 * stays intact while we hold the inode locked. However, vp 1544 * can be NULL here. 1545 */ 1546 vp = ip->vp; 1547 cpu_ccfence(); 1548 1549 /* 1550 * If no vp is associated there is no high-level state to 1551 * deal with and we can scrap the inode immediately. 1552 */ 1553 if (vp == NULL) { 1554 if ((ip->flags & HAMMER2_INODE_DELETING) == 0) { 1555 atomic_set_int(&ip->flags, 1556 HAMMER2_INODE_DELETING); 1557 hammer2_inode_delayed_sideq(ip); 1558 } 1559 return 0; 1560 } 1561 1562 /* 1563 * Because INODE_ISUNLINKED is set with the inode lock 1564 * held, the vnode cannot be ripped up from under us. 1565 * There may still be refs so knote anyone waiting for 1566 * a delete notification. 1567 * 1568 * The vnode is not necessarily ref'd due to the unlinking 1569 * itself, so we have to defer handling to the end of the 1570 * VOP, which will then call hammer2_inode_vprecycle(). 1571 */ 1572 if (vprecyclep) { 1573 vhold(vp); 1574 *vprecyclep = vp; 1575 } 1576 } 1577 1578 /* 1579 * Adjust nlinks and retain the inode on the media for now 1580 */ 1581 hammer2_inode_modify(ip); 1582 if ((int64_t)ip->meta.nlinks > 1) 1583 --ip->meta.nlinks; 1584 else 1585 ip->meta.nlinks = 0; 1586 1587 return 0; 1588 } 1589 1590 /* 1591 * Called at the end of a VOP that removes a file with a vnode that 1592 * we want to try to dispose of quickly due to a file deletion. If 1593 * we don't do this, the vnode can hang around with 0 refs for a very 1594 * long time and prevent reclamation of the underlying file and inode 1595 * (inode remains on-media with nlinks == 0 until the vnode is recycled 1596 * due to random system activity or a umount). 1597 */ 1598 void 1599 hammer2_inode_vprecycle(struct m_vnode *vp) 1600 { 1601 if (vget(vp, LK_EXCLUSIVE) == 0) { 1602 vfinalize(vp); 1603 hammer2_knote(vp, NOTE_DELETE); 1604 vdrop(vp); 1605 vput(vp); 1606 } else { 1607 vdrop(vp); 1608 } 1609 } 1610 1611 1612 /* 1613 * Mark an inode as being modified, meaning that the caller will modify 1614 * ip->meta. 1615 * 1616 * If a vnode is present we set the vnode dirty and the nominal filesystem 1617 * sync will also handle synchronizing the inode meta-data. Unless NOSIDEQ 1618 * we must ensure that the inode is on pmp->sideq. 1619 * 1620 * NOTE: We must always queue the inode to the sideq. This allows H2 to 1621 * shortcut vsyncscan() and flush inodes and their related vnodes 1622 * in a two stages. H2 still calls vfsync() for each vnode. 1623 * 1624 * NOTE: No mtid (modify_tid) is passed into this routine. The caller is 1625 * only modifying the in-memory inode. A modify_tid is synchronized 1626 * later when the inode gets flushed. 1627 * 1628 * NOTE: As an exception to the general rule, the inode MAY be locked 1629 * shared for this particular call. 1630 */ 1631 void 1632 hammer2_inode_modify(hammer2_inode_t *ip) 1633 { 1634 atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED); 1635 if (ip->vp) 1636 vsetisdirty(ip->vp); 1637 if (ip->pmp && (ip->flags & HAMMER2_INODE_NOSIDEQ) == 0) 1638 hammer2_inode_delayed_sideq(ip); 1639 } 1640 1641 /* 1642 * Synchronize the inode's frontend state with the chain state prior 1643 * to any explicit flush of the inode or any strategy write call. This 1644 * does not flush the inode's chain or its sub-topology to media (higher 1645 * level layers are responsible for doing that). 1646 * 1647 * Called with a locked inode inside a normal transaction. 1648 * 1649 * inode must be locked. 1650 */ 1651 int 1652 hammer2_inode_chain_sync(hammer2_inode_t *ip) 1653 { 1654 int error; 1655 1656 error = 0; 1657 if (ip->flags & (HAMMER2_INODE_RESIZED | HAMMER2_INODE_MODIFIED)) { 1658 hammer2_xop_fsync_t *xop; 1659 1660 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1661 xop->clear_directdata = 0; 1662 if (ip->flags & HAMMER2_INODE_RESIZED) { 1663 if ((ip->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) && 1664 ip->meta.size > HAMMER2_EMBEDDED_BYTES) { 1665 ip->meta.op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA; 1666 xop->clear_directdata = 1; 1667 } 1668 xop->osize = ip->osize; 1669 } else { 1670 xop->osize = ip->meta.size; /* safety */ 1671 } 1672 xop->ipflags = ip->flags; 1673 xop->meta = ip->meta; 1674 1675 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED | 1676 HAMMER2_INODE_MODIFIED); 1677 hammer2_xop_start(&xop->head, &hammer2_inode_chain_sync_desc); 1678 error = hammer2_xop_collect(&xop->head, 0); 1679 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1680 if (error == HAMMER2_ERROR_ENOENT) 1681 error = 0; 1682 if (error) { 1683 kprintf("hammer2: unable to fsync inode %p\n", ip); 1684 /* 1685 atomic_set_int(&ip->flags, 1686 xop->ipflags & (HAMMER2_INODE_RESIZED | 1687 HAMMER2_INODE_MODIFIED)); 1688 */ 1689 /* XXX return error somehow? */ 1690 } 1691 } 1692 return error; 1693 } 1694 1695 /* 1696 * When an inode is flagged INODE_CREATING its chains have not actually 1697 * been inserting into the on-media tree yet. 1698 */ 1699 int 1700 hammer2_inode_chain_ins(hammer2_inode_t *ip) 1701 { 1702 int error; 1703 1704 error = 0; 1705 if (ip->flags & HAMMER2_INODE_CREATING) { 1706 hammer2_xop_create_t *xop; 1707 1708 atomic_clear_int(&ip->flags, HAMMER2_INODE_CREATING); 1709 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1710 xop->lhc = ip->meta.inum; 1711 xop->flags = 0; 1712 hammer2_xop_start(&xop->head, &hammer2_inode_create_ins_desc); 1713 error = hammer2_xop_collect(&xop->head, 0); 1714 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1715 if (error == HAMMER2_ERROR_ENOENT) 1716 error = 0; 1717 if (error) { 1718 kprintf("hammer2: backend unable to " 1719 "insert inode %p %ld\n", ip, ip->meta.inum); 1720 /* XXX return error somehow? */ 1721 } 1722 } 1723 return error; 1724 } 1725 1726 /* 1727 * When an inode is flagged INODE_DELETING it has been deleted (no directory 1728 * entry or open refs are left, though as an optimization H2 might leave 1729 * nlinks == 1 to avoid unnecessary block updates). The backend flush then 1730 * needs to actually remove it from the topology. 1731 * 1732 * NOTE: backend flush must still sync and flush the deleted inode to clean 1733 * out related chains. 1734 * 1735 * NOTE: We must clear not only INODE_DELETING, but also INODE_ISUNLINKED 1736 * to prevent the vnode reclaim code from trying to delete it twice. 1737 */ 1738 int 1739 hammer2_inode_chain_des(hammer2_inode_t *ip) 1740 { 1741 int error; 1742 1743 error = 0; 1744 if (ip->flags & HAMMER2_INODE_DELETING) { 1745 hammer2_xop_destroy_t *xop; 1746 1747 atomic_clear_int(&ip->flags, HAMMER2_INODE_DELETING | 1748 HAMMER2_INODE_ISUNLINKED); 1749 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1750 hammer2_xop_start(&xop->head, &hammer2_inode_destroy_desc); 1751 error = hammer2_xop_collect(&xop->head, 0); 1752 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1753 1754 if (error == HAMMER2_ERROR_ENOENT) 1755 error = 0; 1756 if (error) { 1757 kprintf("hammer2: backend unable to " 1758 "delete inode %p %ld\n", ip, ip->meta.inum); 1759 /* XXX return error somehow? */ 1760 } 1761 } 1762 return error; 1763 } 1764 1765 /* 1766 * Flushes the inode's chain and its sub-topology to media. Interlocks 1767 * HAMMER2_INODE_DIRTYDATA by clearing it prior to the flush. Any strategy 1768 * function creating or modifying a chain under this inode will re-set the 1769 * flag. 1770 * 1771 * inode must be locked. 1772 */ 1773 int 1774 hammer2_inode_chain_flush(hammer2_inode_t *ip, int flags) 1775 { 1776 hammer2_xop_fsync_t *xop; 1777 int error; 1778 1779 atomic_clear_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1780 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | flags); 1781 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc); 1782 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_WAITALL); 1783 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1784 if (error == HAMMER2_ERROR_ENOENT) 1785 error = 0; 1786 1787 return error; 1788 } 1789 1790 hammer2_key_t 1791 hammer2_pfs_inode_count(hammer2_pfs_t *pmp) 1792 { 1793 struct hammer2_inode *ip; 1794 hammer2_key_t count = 0; 1795 1796 hammer2_spin_ex(&pmp->inum_spin); 1797 RB_FOREACH(ip, hammer2_inode_tree, &pmp->inum_tree) 1798 count++; 1799 hammer2_spin_unex(&pmp->inum_spin); 1800 1801 return count; 1802 } 1803 1804 int 1805 vflush(struct mount *mp, int rootrefs, int flags) 1806 { 1807 hammer2_pfs_t *pmp = MPTOPMP(mp); 1808 struct hammer2_inode *ip, *tmp; 1809 struct m_vnode *vp; 1810 hammer2_key_t count_before, count_after, count_delta; 1811 1812 hammer2_spin_ex(&pmp->inum_spin); 1813 count_before = 0; 1814 RB_FOREACH(ip, hammer2_inode_tree, &pmp->inum_tree) 1815 count_before++; 1816 1817 RB_FOREACH_SAFE(ip, hammer2_inode_tree, &pmp->inum_tree, tmp) { 1818 vp = ip->vp; 1819 assert(vp); 1820 if (!vp->v_vflushed) { 1821 /* 1822 printf("%s: drop ip=%p inum=%ld refs=%d\n", 1823 __func__, ip, ip->meta.inum, ip->refs); 1824 */ 1825 assert(ip->refs > 1); 1826 hammer2_inode_drop(ip); 1827 vp->v_vflushed = 1; 1828 } 1829 } 1830 1831 count_after = 0; 1832 RB_FOREACH(ip, hammer2_inode_tree, &pmp->inum_tree) 1833 count_after++; 1834 hammer2_spin_unex(&pmp->inum_spin); 1835 1836 printf("%s: total inode %ld -> %ld\n", 1837 __func__, count_before, count_after); 1838 1839 assert(count_before >= count_after); 1840 count_delta = count_before - count_after; 1841 1842 if (count_delta) { 1843 if (hammer2_debug & 0x80000000) 1844 assert(0); 1845 else 1846 printf("%s: %ld inode freed\n", __func__, count_delta); 1847 } 1848 1849 return 0; 1850 } 1851