1 /* 2 * Copyright (c) 2011-2018 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 #include <sys/cdefs.h> 36 #include <sys/param.h> 37 #include <sys/systm.h> 38 #include <sys/types.h> 39 #include <sys/lock.h> 40 #include <sys/uuid.h> 41 #include <sys/vnode.h> 42 43 #include "hammer2.h" 44 45 #define INODE_DEBUG 0 46 47 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp, 48 hammer2_tid_t, meta.inum); 49 50 int 51 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2) 52 { 53 if (ip1->meta.inum < ip2->meta.inum) 54 return(-1); 55 if (ip1->meta.inum > ip2->meta.inum) 56 return(1); 57 return(0); 58 } 59 60 /* 61 * Caller holds pmp->list_spin and the inode should be locked. Merge ip 62 * with the specified depend. 63 * 64 * If the ip is on SYNCQ it stays there and (void *)-1 is returned, indicating 65 * that successive calls must ensure the ip is on a pass2 depend (or they are 66 * all SYNCQ). If the passed-in depend is not NULL and not (void *)-1 then 67 * we can set pass2 on it and return. 68 * 69 * If the ip is not on SYNCQ it is merged with the passed-in depend, creating 70 * a self-depend if necessary, and depend->pass2 is set according 71 * to the PASS2 flag. SIDEQ is set. 72 */ 73 static __noinline 74 hammer2_depend_t * 75 hammer2_inode_setdepend_locked(hammer2_inode_t *ip, hammer2_depend_t *depend) 76 { 77 hammer2_pfs_t *pmp = ip->pmp; 78 hammer2_depend_t *dtmp; 79 hammer2_inode_t *iptmp; 80 81 /* 82 * If ip is SYNCQ its entry is used for the syncq list and it will 83 * no longer be associated with a dependency. Merging this status 84 * with a passed-in depend implies PASS2. 85 */ 86 if (ip->flags & HAMMER2_INODE_SYNCQ) { 87 if (depend == (void *)-1 || 88 depend == NULL) { 89 return ((void *)-1); 90 } 91 depend->pass2 = 1; 92 hammer2_trans_setflags(pmp, HAMMER2_TRANS_RESCAN); 93 94 return depend; 95 } 96 97 /* 98 * If ip is already SIDEQ, merge ip->depend into the passed-in depend. 99 * If it is not, associate the ip with the passed-in depend, creating 100 * a single-entry dependency using depend_static if necessary. 101 * 102 * NOTE: The use of ip->depend_static always requires that the 103 * specific ip containing the structure is part of that 104 * particular depend_static's dependency group. 105 */ 106 if (ip->flags & HAMMER2_INODE_SIDEQ) { 107 /* 108 * Merge ip->depend with the passed-in depend. If the 109 * passed-in depend is not a special case, all ips associated 110 * with ip->depend (including the original ip) must be moved 111 * to the passed-in depend. 112 */ 113 if (depend == NULL) { 114 depend = ip->depend; 115 } else if (depend == (void *)-1) { 116 depend = ip->depend; 117 depend->pass2 = 1; 118 } else if (depend != ip->depend) { 119 #ifdef INVARIANTS 120 int sanitychk = 0; 121 #endif 122 dtmp = ip->depend; 123 while ((iptmp = TAILQ_FIRST(&dtmp->sideq)) != NULL) { 124 #ifdef INVARIANTS 125 if (iptmp == ip) 126 sanitychk = 1; 127 #endif 128 TAILQ_REMOVE(&dtmp->sideq, iptmp, entry); 129 TAILQ_INSERT_TAIL(&depend->sideq, iptmp, entry); 130 iptmp->depend = depend; 131 } 132 KKASSERT(sanitychk == 1); 133 depend->count += dtmp->count; 134 depend->pass2 |= dtmp->pass2; 135 TAILQ_REMOVE(&pmp->depq, dtmp, entry); 136 dtmp->count = 0; 137 dtmp->pass2 = 0; 138 } 139 } else { 140 /* 141 * Add ip to the sideq, creating a self-dependency if 142 * necessary. 143 */ 144 hammer2_inode_ref(ip); 145 atomic_set_int(&ip->flags, HAMMER2_INODE_SIDEQ); 146 if (depend == NULL) { 147 depend = &ip->depend_static; 148 TAILQ_INSERT_TAIL(&pmp->depq, depend, entry); 149 } else if (depend == (void *)-1) { 150 depend = &ip->depend_static; 151 depend->pass2 = 1; 152 TAILQ_INSERT_TAIL(&pmp->depq, depend, entry); 153 } /* else add ip to passed-in depend */ 154 TAILQ_INSERT_TAIL(&depend->sideq, ip, entry); 155 ip->depend = depend; 156 ++depend->count; 157 ++pmp->sideq_count; 158 } 159 160 if (ip->flags & HAMMER2_INODE_SYNCQ_PASS2) 161 depend->pass2 = 1; 162 if (depend->pass2) 163 hammer2_trans_setflags(pmp, HAMMER2_TRANS_RESCAN); 164 165 return depend; 166 } 167 168 /* 169 * Put a solo inode on the SIDEQ (meaning that its dirty). This can also 170 * occur from inode_lock4() and inode_depend(). 171 * 172 * Caller must pass-in a locked inode. 173 */ 174 void 175 hammer2_inode_delayed_sideq(hammer2_inode_t *ip) 176 { 177 hammer2_pfs_t *pmp = ip->pmp; 178 179 /* 180 * Optimize case to avoid pmp spinlock. 181 */ 182 if ((ip->flags & (HAMMER2_INODE_SYNCQ | HAMMER2_INODE_SIDEQ)) == 0) { 183 hammer2_spin_ex(&pmp->list_spin); 184 hammer2_inode_setdepend_locked(ip, NULL); 185 hammer2_spin_unex(&pmp->list_spin); 186 } 187 } 188 189 /* 190 * Lock an inode, with SYNCQ semantics. 191 * 192 * HAMMER2 offers shared and exclusive locks on inodes. Pass a mask of 193 * flags for options: 194 * 195 * - pass HAMMER2_RESOLVE_SHARED if a shared lock is desired. 196 * shared locks are not subject to SYNCQ semantics, exclusive locks 197 * are. 198 * 199 * - pass HAMMER2_RESOLVE_ALWAYS if you need the inode's meta-data. 200 * Most front-end inode locks do. 201 * 202 * - pass HAMMER2_RESOLVE_NEVER if you do not want to require that 203 * the inode data be resolved. This is used by the syncthr because 204 * it can run on an unresolved/out-of-sync cluster, and also by the 205 * vnode reclamation code to avoid unnecessary I/O (particularly when 206 * disposing of hundreds of thousands of cached vnodes). 207 * 208 * This function, along with lock4, has SYNCQ semantics. If the inode being 209 * locked is on the SYNCQ, that is it has been staged by the syncer, we must 210 * block until the operation is complete (even if we can lock the inode). In 211 * order to reduce the stall time, we re-order the inode to the front of the 212 * pmp->syncq prior to blocking. This reordering VERY significantly improves 213 * performance. 214 * 215 * The inode locking function locks the inode itself, resolves any stale 216 * chains in the inode's cluster, and allocates a fresh copy of the 217 * cluster with 1 ref and all the underlying chains locked. 218 * 219 * ip->cluster will be stable while the inode is locked. 220 * 221 * NOTE: We don't combine the inode/chain lock because putting away an 222 * inode would otherwise confuse multiple lock holders of the inode. 223 */ 224 void 225 hammer2_inode_lock(hammer2_inode_t *ip, int how) 226 { 227 hammer2_pfs_t *pmp; 228 229 hammer2_inode_ref(ip); 230 pmp = ip->pmp; 231 232 /* 233 * Inode structure mutex - Shared lock 234 */ 235 if (how & HAMMER2_RESOLVE_SHARED) { 236 hammer2_mtx_sh(&ip->lock); 237 return; 238 } 239 240 /* 241 * Inode structure mutex - Exclusive lock 242 * 243 * An exclusive lock (if not recursive) must wait for inodes on 244 * SYNCQ to flush first, to ensure that meta-data dependencies such 245 * as the nlink count and related directory entries are not split 246 * across flushes. 247 * 248 * If the vnode is locked by the current thread it must be unlocked 249 * across the tsleep() to avoid a deadlock. 250 */ 251 hammer2_mtx_ex(&ip->lock); 252 if (hammer2_mtx_refs(&ip->lock) > 1) 253 return; 254 while ((ip->flags & HAMMER2_INODE_SYNCQ) && pmp) { 255 hammer2_spin_ex(&pmp->list_spin); 256 if (ip->flags & HAMMER2_INODE_SYNCQ) { 257 tsleep_interlock(&ip->flags, 0); 258 atomic_set_int(&ip->flags, HAMMER2_INODE_SYNCQ_WAKEUP); 259 TAILQ_REMOVE(&pmp->syncq, ip, entry); 260 TAILQ_INSERT_HEAD(&pmp->syncq, ip, entry); 261 hammer2_spin_unex(&pmp->list_spin); 262 hammer2_mtx_unlock(&ip->lock); 263 tsleep(&ip->flags, PINTERLOCKED, "h2sync", 0); 264 hammer2_mtx_ex(&ip->lock); 265 continue; 266 } 267 hammer2_spin_unex(&pmp->list_spin); 268 break; 269 } 270 } 271 272 /* 273 * Exclusively lock up to four inodes, in order, with SYNCQ semantics. 274 * ip1 and ip2 must not be NULL. ip3 and ip4 may be NULL, but if ip3 is 275 * NULL then ip4 must also be NULL. 276 * 277 * This creates a dependency between up to four inodes. 278 */ 279 void 280 hammer2_inode_lock4(hammer2_inode_t *ip1, hammer2_inode_t *ip2, 281 hammer2_inode_t *ip3, hammer2_inode_t *ip4) 282 { 283 hammer2_inode_t *ips[4]; 284 hammer2_inode_t *iptmp; 285 hammer2_inode_t *ipslp; 286 hammer2_depend_t *depend; 287 hammer2_pfs_t *pmp; 288 size_t count; 289 size_t i; 290 291 pmp = ip1->pmp; /* may be NULL */ 292 KKASSERT(pmp == ip2->pmp); 293 294 ips[0] = ip1; 295 ips[1] = ip2; 296 if (ip3 == NULL) { 297 count = 2; 298 } else if (ip4 == NULL) { 299 count = 3; 300 ips[2] = ip3; 301 KKASSERT(pmp == ip3->pmp); 302 } else { 303 count = 4; 304 ips[2] = ip3; 305 ips[3] = ip4; 306 KKASSERT(pmp == ip3->pmp); 307 KKASSERT(pmp == ip4->pmp); 308 } 309 310 for (i = 0; i < count; ++i) 311 hammer2_inode_ref(ips[i]); 312 313 restart: 314 /* 315 * Lock the inodes in order 316 */ 317 for (i = 0; i < count; ++i) { 318 hammer2_mtx_ex(&ips[i]->lock); 319 } 320 321 /* 322 * Associate dependencies, record the first inode found on SYNCQ 323 * (operation is allowed to proceed for inodes on PASS2) for our 324 * sleep operation, this inode is theoretically the last one sync'd 325 * in the sequence. 326 * 327 * All inodes found on SYNCQ are moved to the head of the syncq 328 * to reduce stalls. 329 */ 330 hammer2_spin_ex(&pmp->list_spin); 331 depend = NULL; 332 ipslp = NULL; 333 for (i = 0; i < count; ++i) { 334 iptmp = ips[i]; 335 depend = hammer2_inode_setdepend_locked(iptmp, depend); 336 if (iptmp->flags & HAMMER2_INODE_SYNCQ) { 337 TAILQ_REMOVE(&pmp->syncq, iptmp, entry); 338 TAILQ_INSERT_HEAD(&pmp->syncq, iptmp, entry); 339 if (ipslp == NULL) 340 ipslp = iptmp; 341 } 342 } 343 hammer2_spin_unex(&pmp->list_spin); 344 345 /* 346 * Block and retry if any of the inodes are on SYNCQ. It is 347 * important that we allow the operation to proceed in the 348 * PASS2 case, to avoid deadlocking against the vnode. 349 */ 350 if (ipslp) { 351 for (i = 0; i < count; ++i) 352 hammer2_mtx_unlock(&ips[i]->lock); 353 tsleep(&ipslp->flags, 0, "h2sync", 2); 354 goto restart; 355 } 356 } 357 358 /* 359 * Release an inode lock. If another thread is blocked on SYNCQ_WAKEUP 360 * we wake them up. 361 */ 362 void 363 hammer2_inode_unlock(hammer2_inode_t *ip) 364 { 365 if (ip->flags & HAMMER2_INODE_SYNCQ_WAKEUP) { 366 atomic_clear_int(&ip->flags, HAMMER2_INODE_SYNCQ_WAKEUP); 367 hammer2_mtx_unlock(&ip->lock); 368 wakeup(&ip->flags); 369 } else { 370 hammer2_mtx_unlock(&ip->lock); 371 } 372 hammer2_inode_drop(ip); 373 } 374 375 /* 376 * If either ip1 or ip2 have been tapped by the syncer, make sure that both 377 * are. This ensure that dependencies (e.g. dirent-v-inode) are synced 378 * together. For dirent-v-inode depends, pass the dirent as ip1. 379 * 380 * If neither ip1 or ip2 have been tapped by the syncer, merge them into a 381 * single dependency. Dependencies are entered into pmp->depq. This 382 * effectively flags the inodes SIDEQ. 383 * 384 * Both ip1 and ip2 must be locked by the caller. This also ensures 385 * that we can't race the end of the syncer's queue run. 386 */ 387 void 388 hammer2_inode_depend(hammer2_inode_t *ip1, hammer2_inode_t *ip2) 389 { 390 hammer2_pfs_t *pmp; 391 hammer2_depend_t *depend; 392 393 pmp = ip1->pmp; 394 hammer2_spin_ex(&pmp->list_spin); 395 depend = hammer2_inode_setdepend_locked(ip1, NULL); 396 depend = hammer2_inode_setdepend_locked(ip2, depend); 397 hammer2_spin_unex(&pmp->list_spin); 398 } 399 400 /* 401 * Select a chain out of an inode's cluster and lock it. 402 * 403 * The inode does not have to be locked. 404 */ 405 hammer2_chain_t * 406 hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how) 407 { 408 hammer2_chain_t *chain; 409 hammer2_cluster_t *cluster; 410 411 hammer2_spin_sh(&ip->cluster_spin); 412 cluster = &ip->cluster; 413 if (clindex >= cluster->nchains) 414 chain = NULL; 415 else 416 chain = cluster->array[clindex].chain; 417 if (chain) { 418 hammer2_chain_ref(chain); 419 hammer2_spin_unsh(&ip->cluster_spin); 420 hammer2_chain_lock(chain, how); 421 } else { 422 hammer2_spin_unsh(&ip->cluster_spin); 423 } 424 return chain; 425 } 426 427 hammer2_chain_t * 428 hammer2_inode_chain_and_parent(hammer2_inode_t *ip, int clindex, 429 hammer2_chain_t **parentp, int how) 430 { 431 hammer2_chain_t *chain; 432 hammer2_chain_t *parent; 433 434 for (;;) { 435 hammer2_spin_sh(&ip->cluster_spin); 436 if (clindex >= ip->cluster.nchains) 437 chain = NULL; 438 else 439 chain = ip->cluster.array[clindex].chain; 440 if (chain) { 441 hammer2_chain_ref(chain); 442 hammer2_spin_unsh(&ip->cluster_spin); 443 hammer2_chain_lock(chain, how); 444 } else { 445 hammer2_spin_unsh(&ip->cluster_spin); 446 } 447 448 /* 449 * Get parent, lock order must be (parent, chain). 450 */ 451 parent = chain->parent; 452 if (parent) { 453 hammer2_chain_ref(parent); 454 hammer2_chain_unlock(chain); 455 hammer2_chain_lock(parent, how); 456 hammer2_chain_lock(chain, how); 457 } 458 if (ip->cluster.array[clindex].chain == chain && 459 chain->parent == parent) { 460 break; 461 } 462 463 /* 464 * Retry 465 */ 466 hammer2_chain_unlock(chain); 467 hammer2_chain_drop(chain); 468 if (parent) { 469 hammer2_chain_unlock(parent); 470 hammer2_chain_drop(parent); 471 } 472 } 473 *parentp = parent; 474 475 return chain; 476 } 477 478 /* 479 * Temporarily release a lock held shared or exclusive. Caller must 480 * hold the lock shared or exclusive on call and lock will be released 481 * on return. 482 * 483 * Restore a lock that was temporarily released. 484 */ 485 hammer2_mtx_state_t 486 hammer2_inode_lock_temp_release(hammer2_inode_t *ip) 487 { 488 return hammer2_mtx_temp_release(&ip->lock); 489 } 490 491 void 492 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, hammer2_mtx_state_t ostate) 493 { 494 hammer2_mtx_temp_restore(&ip->lock, ostate); 495 } 496 497 /* 498 * Upgrade a shared inode lock to exclusive and return. If the inode lock 499 * is already held exclusively this is a NOP. 500 * 501 * The caller MUST hold the inode lock either shared or exclusive on call 502 * and will own the lock exclusively on return. 503 * 504 * Returns non-zero if the lock was already exclusive prior to the upgrade. 505 */ 506 int 507 hammer2_inode_lock_upgrade(hammer2_inode_t *ip) 508 { 509 int wasexclusive; 510 511 if (mtx_islocked_ex(&ip->lock)) { 512 wasexclusive = 1; 513 } else { 514 hammer2_mtx_unlock(&ip->lock); 515 hammer2_mtx_ex(&ip->lock); 516 wasexclusive = 0; 517 } 518 return wasexclusive; 519 } 520 521 /* 522 * Downgrade an inode lock from exclusive to shared only if the inode 523 * lock was previously shared. If the inode lock was previously exclusive, 524 * this is a NOP. 525 */ 526 void 527 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int wasexclusive) 528 { 529 if (wasexclusive == 0) 530 hammer2_mtx_downgrade(&ip->lock); 531 } 532 533 /* 534 * Lookup an inode by inode number 535 */ 536 hammer2_inode_t * 537 hammer2_inode_lookup(hammer2_pfs_t *pmp, hammer2_tid_t inum) 538 { 539 hammer2_inode_t *ip; 540 541 KKASSERT(pmp); 542 if (pmp->spmp_hmp) { 543 ip = NULL; 544 } else { 545 hammer2_spin_ex(&pmp->inum_spin); 546 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum); 547 if (ip) 548 hammer2_inode_ref(ip); 549 hammer2_spin_unex(&pmp->inum_spin); 550 } 551 return(ip); 552 } 553 554 /* 555 * Adding a ref to an inode is only legal if the inode already has at least 556 * one ref. 557 * 558 * (can be called with spinlock held) 559 */ 560 void 561 hammer2_inode_ref(hammer2_inode_t *ip) 562 { 563 atomic_add_int(&ip->refs, 1); 564 if (hammer2_debug & 0x80000) { 565 kprintf("INODE+1 %p (%d->%d)\n", ip, ip->refs - 1, ip->refs); 566 print_backtrace(8); 567 } 568 } 569 570 /* 571 * Drop an inode reference, freeing the inode when the last reference goes 572 * away. 573 */ 574 void 575 hammer2_inode_drop(hammer2_inode_t *ip) 576 { 577 hammer2_pfs_t *pmp; 578 u_int refs; 579 580 while (ip) { 581 if (hammer2_debug & 0x80000) { 582 kprintf("INODE-1 %p (%d->%d)\n", 583 ip, ip->refs, ip->refs - 1); 584 print_backtrace(8); 585 } 586 refs = ip->refs; 587 cpu_ccfence(); 588 if (refs == 1) { 589 /* 590 * Transition to zero, must interlock with 591 * the inode inumber lookup tree (if applicable). 592 * It should not be possible for anyone to race 593 * the transition to 0. 594 */ 595 pmp = ip->pmp; 596 KKASSERT(pmp); 597 hammer2_spin_ex(&pmp->inum_spin); 598 599 if (atomic_cmpset_int(&ip->refs, 1, 0)) { 600 KKASSERT(hammer2_mtx_refs(&ip->lock) == 0); 601 if (ip->flags & HAMMER2_INODE_ONRBTREE) { 602 atomic_clear_int(&ip->flags, 603 HAMMER2_INODE_ONRBTREE); 604 RB_REMOVE(hammer2_inode_tree, 605 &pmp->inum_tree, ip); 606 --pmp->inum_count; 607 } 608 hammer2_spin_unex(&pmp->inum_spin); 609 610 ip->pmp = NULL; 611 612 /* 613 * Cleaning out ip->cluster isn't entirely 614 * trivial. 615 */ 616 hammer2_inode_repoint(ip, NULL); 617 618 kfree_obj(ip, pmp->minode); 619 atomic_add_long(&pmp->inmem_inodes, -1); 620 ip = NULL; /* will terminate loop */ 621 } else { 622 hammer2_spin_unex(&ip->pmp->inum_spin); 623 } 624 } else { 625 /* 626 * Non zero transition 627 */ 628 if (atomic_cmpset_int(&ip->refs, refs, refs - 1)) 629 break; 630 } 631 } 632 } 633 634 /* 635 * Get the vnode associated with the given inode, allocating the vnode if 636 * necessary. The vnode will be returned exclusively locked. 637 * 638 * *errorp is set to a UNIX error, not a HAMMER2 error. 639 * 640 * The caller must lock the inode (shared or exclusive). 641 * 642 * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim 643 * races. 644 */ 645 struct vnode * 646 hammer2_igetv(hammer2_inode_t *ip, int *errorp) 647 { 648 hammer2_pfs_t *pmp; 649 struct vnode *vp; 650 651 pmp = ip->pmp; 652 KKASSERT(pmp != NULL); 653 *errorp = 0; 654 655 for (;;) { 656 /* 657 * Attempt to reuse an existing vnode assignment. It is 658 * possible to race a reclaim so the vget() may fail. The 659 * inode must be unlocked during the vget() to avoid a 660 * deadlock against a reclaim. 661 */ 662 int wasexclusive; 663 664 vp = ip->vp; 665 if (vp) { 666 /* 667 * Inode must be unlocked during the vget() to avoid 668 * possible deadlocks, but leave the ip ref intact. 669 * 670 * vnode is held to prevent destruction during the 671 * vget(). The vget() can still fail if we lost 672 * a reclaim race on the vnode. 673 */ 674 hammer2_mtx_state_t ostate; 675 676 vhold(vp); 677 ostate = hammer2_inode_lock_temp_release(ip); 678 if (vget(vp, LK_EXCLUSIVE)) { 679 vdrop(vp); 680 hammer2_inode_lock_temp_restore(ip, ostate); 681 continue; 682 } 683 hammer2_inode_lock_temp_restore(ip, ostate); 684 vdrop(vp); 685 /* vp still locked and ref from vget */ 686 if (ip->vp != vp) { 687 kprintf("hammer2: igetv race %p/%p\n", 688 ip->vp, vp); 689 vput(vp); 690 continue; 691 } 692 *errorp = 0; 693 break; 694 } 695 696 /* 697 * No vnode exists, allocate a new vnode. Beware of 698 * allocation races. This function will return an 699 * exclusively locked and referenced vnode. 700 */ 701 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0); 702 if (*errorp) { 703 kprintf("hammer2: igetv getnewvnode failed %d\n", 704 *errorp); 705 vp = NULL; 706 break; 707 } 708 709 /* 710 * Lock the inode and check for an allocation race. 711 */ 712 wasexclusive = hammer2_inode_lock_upgrade(ip); 713 if (ip->vp != NULL) { 714 vp->v_type = VBAD; 715 vx_put(vp); 716 hammer2_inode_lock_downgrade(ip, wasexclusive); 717 continue; 718 } 719 720 switch (ip->meta.type) { 721 case HAMMER2_OBJTYPE_DIRECTORY: 722 vp->v_type = VDIR; 723 break; 724 case HAMMER2_OBJTYPE_REGFILE: 725 /* 726 * Regular file must use buffer cache I/O 727 * (VKVABIO cpu sync semantics supported) 728 */ 729 vp->v_type = VREG; 730 vsetflags(vp, VKVABIO); 731 vinitvmio(vp, ip->meta.size, 732 HAMMER2_LBUFSIZE, 733 (int)ip->meta.size & HAMMER2_LBUFMASK); 734 break; 735 case HAMMER2_OBJTYPE_SOFTLINK: 736 /* 737 * XXX for now we are using the generic file_read 738 * and file_write code so we need a buffer cache 739 * association. 740 * 741 * (VKVABIO cpu sync semantics supported) 742 */ 743 vp->v_type = VLNK; 744 vsetflags(vp, VKVABIO); 745 vinitvmio(vp, ip->meta.size, 746 HAMMER2_LBUFSIZE, 747 (int)ip->meta.size & HAMMER2_LBUFMASK); 748 break; 749 case HAMMER2_OBJTYPE_CDEV: 750 vp->v_type = VCHR; 751 /* fall through */ 752 case HAMMER2_OBJTYPE_BDEV: 753 vp->v_ops = &pmp->mp->mnt_vn_spec_ops; 754 if (ip->meta.type != HAMMER2_OBJTYPE_CDEV) 755 vp->v_type = VBLK; 756 addaliasu(vp, 757 ip->meta.rmajor, 758 ip->meta.rminor); 759 break; 760 case HAMMER2_OBJTYPE_FIFO: 761 vp->v_type = VFIFO; 762 vp->v_ops = &pmp->mp->mnt_vn_fifo_ops; 763 break; 764 case HAMMER2_OBJTYPE_SOCKET: 765 vp->v_type = VSOCK; 766 break; 767 default: 768 panic("hammer2: unhandled objtype %d", 769 ip->meta.type); 770 break; 771 } 772 773 if (ip == pmp->iroot) 774 vsetflags(vp, VROOT); 775 776 vp->v_data = ip; 777 ip->vp = vp; 778 hammer2_inode_ref(ip); /* vp association */ 779 hammer2_inode_lock_downgrade(ip, wasexclusive); 780 vx_downgrade(vp); 781 break; 782 } 783 784 /* 785 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0. 786 */ 787 if (hammer2_debug & 0x0002) { 788 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n", 789 vp, vp->v_refcnt, vp->v_auxrefs); 790 } 791 return (vp); 792 } 793 794 /* 795 * XXX this API needs a rewrite. It needs to be split into a 796 * hammer2_inode_alloc() and hammer2_inode_build() to allow us to get 797 * rid of the inode/chain lock reversal fudge. 798 * 799 * Returns the inode associated with the passed-in cluster, allocating a new 800 * hammer2_inode structure if necessary, then synchronizing it to the passed 801 * xop cluster. When synchronizing, if idx >= 0, only cluster index (idx) 802 * is synchronized. Otherwise the whole cluster is synchronized. inum will 803 * be extracted from the passed-in xop and the inum argument will be ignored. 804 * 805 * If xop is passed as NULL then a new hammer2_inode is allocated with the 806 * specified inum, and returned. For normal inodes, the inode will be 807 * indexed in memory and if it already exists the existing ip will be 808 * returned instead of allocating a new one. The superroot and PFS inodes 809 * are not indexed in memory. 810 * 811 * The passed-in cluster must be locked and will remain locked on return. 812 * The returned inode will be locked and the caller may dispose of both 813 * via hammer2_inode_unlock() + hammer2_inode_drop(). However, if the caller 814 * needs to resolve a hardlink it must ref/unlock/relock/drop the inode. 815 * 816 * The hammer2_inode structure regulates the interface between the high level 817 * kernel VNOPS API and the filesystem backend (the chains). 818 * 819 * On return the inode is locked with the supplied cluster. 820 */ 821 hammer2_inode_t * 822 hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_xop_head_t *xop, 823 hammer2_tid_t inum, int idx) 824 { 825 hammer2_inode_t *nip; 826 const hammer2_inode_data_t *iptmp; 827 const hammer2_inode_data_t *nipdata; 828 829 KKASSERT(xop == NULL || 830 hammer2_cluster_type(&xop->cluster) == 831 HAMMER2_BREF_TYPE_INODE); 832 KKASSERT(pmp); 833 834 /* 835 * Interlocked lookup/ref of the inode. This code is only needed 836 * when looking up inodes with nlinks != 0 (TODO: optimize out 837 * otherwise and test for duplicates). 838 * 839 * Cluster can be NULL during the initial pfs allocation. 840 */ 841 if (xop) { 842 iptmp = &hammer2_xop_gdata(xop)->ipdata; 843 inum = iptmp->meta.inum; 844 hammer2_xop_pdata(xop); 845 } 846 again: 847 nip = hammer2_inode_lookup(pmp, inum); 848 if (nip) { 849 /* 850 * We may have to unhold the cluster to avoid a deadlock 851 * against vnlru (and possibly other XOPs). 852 */ 853 if (xop) { 854 if (hammer2_mtx_ex_try(&nip->lock) != 0) { 855 hammer2_cluster_unhold(&xop->cluster); 856 hammer2_mtx_ex(&nip->lock); 857 hammer2_cluster_rehold(&xop->cluster); 858 } 859 } else { 860 hammer2_mtx_ex(&nip->lock); 861 } 862 863 /* 864 * Handle SMP race (not applicable to the super-root spmp 865 * which can't index inodes due to duplicative inode numbers). 866 */ 867 if (pmp->spmp_hmp == NULL && 868 (nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { 869 hammer2_mtx_unlock(&nip->lock); 870 hammer2_inode_drop(nip); 871 goto again; 872 } 873 if (xop) { 874 if (idx >= 0) 875 hammer2_inode_repoint_one(nip, &xop->cluster, 876 idx); 877 else 878 hammer2_inode_repoint(nip, &xop->cluster); 879 } 880 return nip; 881 } 882 883 /* 884 * We couldn't find the inode number, create a new inode and try to 885 * insert it, handle insertion races. 886 */ 887 nip = kmalloc_obj(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO); 888 hammer2_spin_init(&nip->cluster_spin, "h2clspin"); 889 atomic_add_long(&pmp->inmem_inodes, 1); 890 891 /* 892 * Initialize nip's cluster. A cluster is provided for normal 893 * inodes but typically not for the super-root or PFS inodes. 894 */ 895 { 896 hammer2_inode_t *nnip = nip; 897 nip->ihash = (int)hammer2_icrc32(&nnip, sizeof(nnip)); 898 } 899 900 nip->cluster.refs = 1; 901 nip->cluster.pmp = pmp; 902 nip->cluster.flags |= HAMMER2_CLUSTER_INODE; 903 if (xop) { 904 nipdata = &hammer2_xop_gdata(xop)->ipdata; 905 nip->meta = nipdata->meta; 906 hammer2_xop_pdata(xop); 907 hammer2_inode_repoint(nip, &xop->cluster); 908 } else { 909 nip->meta.inum = inum; /* PFS inum is always 1 XXX */ 910 /* mtime will be updated when a cluster is available */ 911 } 912 913 nip->pmp = pmp; 914 915 /* 916 * ref and lock on nip gives it state compatible to after a 917 * hammer2_inode_lock() call. 918 */ 919 nip->refs = 1; 920 hammer2_mtx_init(&nip->lock, "h2inode"); 921 hammer2_mtx_init(&nip->truncate_lock, "h2trunc"); 922 hammer2_mtx_ex(&nip->lock); 923 TAILQ_INIT(&nip->depend_static.sideq); 924 /* combination of thread lock and chain lock == inode lock */ 925 926 /* 927 * Attempt to add the inode. If it fails we raced another inode 928 * get. Undo all the work and try again. 929 */ 930 if (pmp->spmp_hmp == NULL) { 931 hammer2_spin_ex(&pmp->inum_spin); 932 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) { 933 hammer2_spin_unex(&pmp->inum_spin); 934 hammer2_mtx_unlock(&nip->lock); 935 hammer2_inode_drop(nip); 936 goto again; 937 } 938 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE); 939 ++pmp->inum_count; 940 hammer2_spin_unex(&pmp->inum_spin); 941 } 942 return (nip); 943 } 944 945 /* 946 * Create a PFS inode under the superroot. This function will create the 947 * inode, its media chains, and also insert it into the media. 948 * 949 * Caller must be in a flush transaction because we are inserting the inode 950 * onto the media. 951 */ 952 hammer2_inode_t * 953 hammer2_inode_create_pfs(hammer2_pfs_t *spmp, 954 const char *name, size_t name_len, 955 int *errorp) 956 { 957 hammer2_xop_create_t *xop; 958 hammer2_inode_t *pip; 959 hammer2_inode_t *nip; 960 int error; 961 uint8_t pip_comp_algo; 962 uint8_t pip_check_algo; 963 hammer2_tid_t pip_inum; 964 hammer2_key_t lhc; 965 966 pip = spmp->iroot; 967 nip = NULL; 968 969 lhc = hammer2_dirhash(name, name_len); 970 *errorp = 0; 971 972 /* 973 * Locate the inode or indirect block to create the new 974 * entry in. At the same time check for key collisions 975 * and iterate until we don't get one. 976 * 977 * Lock the directory exclusively for now to guarantee that 978 * we can find an unused lhc for the name. Due to collisions, 979 * two different creates can end up with the same lhc so we 980 * cannot depend on the OS to prevent the collision. 981 */ 982 hammer2_inode_lock(pip, 0); 983 984 pip_comp_algo = pip->meta.comp_algo; 985 pip_check_algo = pip->meta.check_algo; 986 pip_inum = (pip == pip->pmp->iroot) ? 1 : pip->meta.inum; 987 988 /* 989 * Locate an unused key in the collision space. 990 */ 991 { 992 hammer2_xop_scanlhc_t *sxop; 993 hammer2_key_t lhcbase; 994 995 lhcbase = lhc; 996 sxop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 997 sxop->lhc = lhc; 998 hammer2_xop_start(&sxop->head, &hammer2_scanlhc_desc); 999 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) { 1000 if (lhc != sxop->head.cluster.focus->bref.key) 1001 break; 1002 ++lhc; 1003 } 1004 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP); 1005 1006 if (error) { 1007 if (error != HAMMER2_ERROR_ENOENT) 1008 goto done2; 1009 ++lhc; 1010 error = 0; 1011 } 1012 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) { 1013 error = HAMMER2_ERROR_ENOSPC; 1014 goto done2; 1015 } 1016 } 1017 1018 /* 1019 * Create the inode with the lhc as the key. 1020 */ 1021 xop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 1022 xop->lhc = lhc; 1023 xop->flags = HAMMER2_INSERT_PFSROOT; 1024 bzero(&xop->meta, sizeof(xop->meta)); 1025 1026 xop->meta.type = HAMMER2_OBJTYPE_DIRECTORY; 1027 xop->meta.inum = 1; 1028 xop->meta.iparent = pip_inum; 1029 1030 /* Inherit parent's inode compression mode. */ 1031 xop->meta.comp_algo = pip_comp_algo; 1032 xop->meta.check_algo = pip_check_algo; 1033 xop->meta.version = HAMMER2_INODE_VERSION_ONE; 1034 hammer2_update_time(&xop->meta.ctime); 1035 xop->meta.mtime = xop->meta.ctime; 1036 xop->meta.mode = 0755; 1037 xop->meta.nlinks = 1; 1038 1039 /* 1040 * Regular files and softlinks allow a small amount of data to be 1041 * directly embedded in the inode. This flag will be cleared if 1042 * the size is extended past the embedded limit. 1043 */ 1044 if (xop->meta.type == HAMMER2_OBJTYPE_REGFILE || 1045 xop->meta.type == HAMMER2_OBJTYPE_SOFTLINK) { 1046 xop->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA; 1047 } 1048 hammer2_xop_setname(&xop->head, name, name_len); 1049 xop->meta.name_len = name_len; 1050 xop->meta.name_key = lhc; 1051 KKASSERT(name_len < HAMMER2_INODE_MAXNAME); 1052 1053 hammer2_xop_start(&xop->head, &hammer2_inode_create_desc); 1054 1055 error = hammer2_xop_collect(&xop->head, 0); 1056 #if INODE_DEBUG 1057 kprintf("CREATE INODE %*.*s\n", 1058 (int)name_len, (int)name_len, name); 1059 #endif 1060 1061 if (error) { 1062 *errorp = error; 1063 goto done; 1064 } 1065 1066 /* 1067 * Set up the new inode if not a hardlink pointer. 1068 * 1069 * NOTE: *_get() integrates chain's lock into the inode lock. 1070 * 1071 * NOTE: Only one new inode can currently be created per 1072 * transaction. If the need arises we can adjust 1073 * hammer2_trans_init() to allow more. 1074 * 1075 * NOTE: nipdata will have chain's blockset data. 1076 */ 1077 nip = hammer2_inode_get(pip->pmp, &xop->head, -1, -1); 1078 nip->comp_heuristic = 0; 1079 done: 1080 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1081 done2: 1082 hammer2_inode_unlock(pip); 1083 1084 return (nip); 1085 } 1086 1087 /* 1088 * Create a new, normal inode. This function will create the inode, 1089 * the media chains, but will not insert the chains onto the media topology 1090 * (doing so would require a flush transaction and cause long stalls). 1091 * 1092 * Caller must be in a normal transaction. 1093 */ 1094 hammer2_inode_t * 1095 hammer2_inode_create_normal(hammer2_inode_t *pip, 1096 struct vattr *vap, struct ucred *cred, 1097 hammer2_key_t inum, int *errorp) 1098 { 1099 hammer2_xop_create_t *xop; 1100 hammer2_inode_t *dip; 1101 hammer2_inode_t *nip; 1102 int error; 1103 uid_t xuid; 1104 uuid_t pip_uid; 1105 uuid_t pip_gid; 1106 uint32_t pip_mode; 1107 uint8_t pip_comp_algo; 1108 uint8_t pip_check_algo; 1109 hammer2_tid_t pip_inum; 1110 1111 dip = pip->pmp->iroot; 1112 KKASSERT(dip != NULL); 1113 1114 *errorp = 0; 1115 1116 /*hammer2_inode_lock(dip, 0);*/ 1117 1118 pip_uid = pip->meta.uid; 1119 pip_gid = pip->meta.gid; 1120 pip_mode = pip->meta.mode; 1121 pip_comp_algo = pip->meta.comp_algo; 1122 pip_check_algo = pip->meta.check_algo; 1123 pip_inum = (pip == pip->pmp->iroot) ? 1 : pip->meta.inum; 1124 1125 /* 1126 * Create the in-memory hammer2_inode structure for the specified 1127 * inode. 1128 */ 1129 nip = hammer2_inode_get(dip->pmp, NULL, inum, -1); 1130 nip->comp_heuristic = 0; 1131 KKASSERT((nip->flags & HAMMER2_INODE_CREATING) == 0 && 1132 nip->cluster.nchains == 0); 1133 atomic_set_int(&nip->flags, HAMMER2_INODE_CREATING); 1134 1135 /* 1136 * Setup the inode meta-data 1137 */ 1138 nip->meta.type = hammer2_get_obj_type(vap->va_type); 1139 1140 switch (nip->meta.type) { 1141 case HAMMER2_OBJTYPE_CDEV: 1142 case HAMMER2_OBJTYPE_BDEV: 1143 nip->meta.rmajor = vap->va_rmajor; 1144 nip->meta.rminor = vap->va_rminor; 1145 break; 1146 default: 1147 break; 1148 } 1149 1150 KKASSERT(nip->meta.inum == inum); 1151 nip->meta.iparent = pip_inum; 1152 1153 /* Inherit parent's inode compression mode. */ 1154 nip->meta.comp_algo = pip_comp_algo; 1155 nip->meta.check_algo = pip_check_algo; 1156 nip->meta.version = HAMMER2_INODE_VERSION_ONE; 1157 hammer2_update_time(&nip->meta.ctime); 1158 nip->meta.mtime = nip->meta.ctime; 1159 nip->meta.mode = vap->va_mode; 1160 nip->meta.nlinks = 1; 1161 1162 xuid = hammer2_to_unix_xid(&pip_uid); 1163 xuid = vop_helper_create_uid(dip->pmp->mp, pip_mode, 1164 xuid, cred, 1165 &vap->va_mode); 1166 if (vap->va_vaflags & VA_UID_UUID_VALID) 1167 nip->meta.uid = vap->va_uid_uuid; 1168 else if (vap->va_uid != (uid_t)VNOVAL) 1169 hammer2_guid_to_uuid(&nip->meta.uid, vap->va_uid); 1170 else 1171 hammer2_guid_to_uuid(&nip->meta.uid, xuid); 1172 1173 if (vap->va_vaflags & VA_GID_UUID_VALID) 1174 nip->meta.gid = vap->va_gid_uuid; 1175 else if (vap->va_gid != (gid_t)VNOVAL) 1176 hammer2_guid_to_uuid(&nip->meta.gid, vap->va_gid); 1177 else 1178 nip->meta.gid = pip_gid; 1179 1180 /* 1181 * Regular files and softlinks allow a small amount of data to be 1182 * directly embedded in the inode. This flag will be cleared if 1183 * the size is extended past the embedded limit. 1184 */ 1185 if (nip->meta.type == HAMMER2_OBJTYPE_REGFILE || 1186 nip->meta.type == HAMMER2_OBJTYPE_SOFTLINK) { 1187 nip->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA; 1188 } 1189 1190 /* 1191 * Create the inode using (inum) as the key. Pass pip for 1192 * method inheritance. 1193 */ 1194 xop = hammer2_xop_alloc(pip, HAMMER2_XOP_MODIFYING); 1195 xop->lhc = inum; 1196 xop->flags = 0; 1197 xop->meta = nip->meta; 1198 KKASSERT(vap); 1199 1200 xop->meta.name_len = hammer2_xop_setname_inum(&xop->head, inum); 1201 xop->meta.name_key = inum; 1202 nip->meta.name_len = xop->meta.name_len; 1203 nip->meta.name_key = xop->meta.name_key; 1204 hammer2_inode_modify(nip); 1205 1206 /* 1207 * Create the inode media chains but leave them detached. We are 1208 * not in a flush transaction so we can't mess with media topology 1209 * above normal inodes (i.e. the index of the inodes themselves). 1210 * 1211 * We've already set the INODE_CREATING flag. The inode's media 1212 * chains will be inserted onto the media topology on the next 1213 * filesystem sync. 1214 */ 1215 hammer2_xop_start(&xop->head, &hammer2_inode_create_det_desc); 1216 1217 error = hammer2_xop_collect(&xop->head, 0); 1218 #if INODE_DEBUG 1219 kprintf("create inode type %d error %d\n", nip->meta.type, error); 1220 #endif 1221 1222 if (error) { 1223 *errorp = error; 1224 goto done; 1225 } 1226 1227 /* 1228 * Associate the media chains created by the backend with the 1229 * frontend inode. 1230 */ 1231 hammer2_inode_repoint(nip, &xop->head.cluster); 1232 done: 1233 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1234 /*hammer2_inode_unlock(dip);*/ 1235 1236 return (nip); 1237 } 1238 1239 /* 1240 * Create a directory entry under dip with the specified name, inode number, 1241 * and OBJTYPE (type). 1242 * 1243 * This returns a UNIX errno code, not a HAMMER2_ERROR_* code. 1244 * 1245 * Caller must hold dip locked. 1246 */ 1247 int 1248 hammer2_dirent_create(hammer2_inode_t *dip, const char *name, size_t name_len, 1249 hammer2_key_t inum, uint8_t type) 1250 { 1251 hammer2_xop_mkdirent_t *xop; 1252 hammer2_key_t lhc; 1253 int error; 1254 1255 lhc = 0; 1256 error = 0; 1257 1258 KKASSERT(name != NULL); 1259 lhc = hammer2_dirhash(name, name_len); 1260 1261 /* 1262 * Locate the inode or indirect block to create the new 1263 * entry in. At the same time check for key collisions 1264 * and iterate until we don't get one. 1265 * 1266 * Lock the directory exclusively for now to guarantee that 1267 * we can find an unused lhc for the name. Due to collisions, 1268 * two different creates can end up with the same lhc so we 1269 * cannot depend on the OS to prevent the collision. 1270 */ 1271 hammer2_inode_modify(dip); 1272 1273 /* 1274 * If name specified, locate an unused key in the collision space. 1275 * Otherwise use the passed-in lhc directly. 1276 */ 1277 { 1278 hammer2_xop_scanlhc_t *sxop; 1279 hammer2_key_t lhcbase; 1280 1281 lhcbase = lhc; 1282 sxop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING); 1283 sxop->lhc = lhc; 1284 hammer2_xop_start(&sxop->head, &hammer2_scanlhc_desc); 1285 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) { 1286 if (lhc != sxop->head.cluster.focus->bref.key) 1287 break; 1288 ++lhc; 1289 } 1290 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP); 1291 1292 if (error) { 1293 if (error != HAMMER2_ERROR_ENOENT) 1294 goto done2; 1295 ++lhc; 1296 error = 0; 1297 } 1298 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) { 1299 error = HAMMER2_ERROR_ENOSPC; 1300 goto done2; 1301 } 1302 } 1303 1304 /* 1305 * Create the directory entry with the lhc as the key. 1306 */ 1307 xop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING); 1308 xop->lhc = lhc; 1309 bzero(&xop->dirent, sizeof(xop->dirent)); 1310 xop->dirent.inum = inum; 1311 xop->dirent.type = type; 1312 xop->dirent.namlen = name_len; 1313 1314 KKASSERT(name_len < HAMMER2_INODE_MAXNAME); 1315 hammer2_xop_setname(&xop->head, name, name_len); 1316 1317 hammer2_xop_start(&xop->head, &hammer2_inode_mkdirent_desc); 1318 1319 error = hammer2_xop_collect(&xop->head, 0); 1320 1321 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1322 done2: 1323 error = hammer2_error_to_errno(error); 1324 1325 return error; 1326 } 1327 1328 /* 1329 * Repoint ip->cluster's chains to cluster's chains and fixup the default 1330 * focus. All items, valid or invalid, are repointed. hammer2_xop_start() 1331 * filters out invalid or non-matching elements. 1332 * 1333 * Caller must hold the inode and cluster exclusive locked, if not NULL, 1334 * must also be locked. 1335 * 1336 * Cluster may be NULL to clean out any chains in ip->cluster. 1337 */ 1338 void 1339 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_cluster_t *cluster) 1340 { 1341 hammer2_chain_t *dropch[HAMMER2_MAXCLUSTER]; 1342 hammer2_chain_t *ochain; 1343 hammer2_chain_t *nchain; 1344 int i; 1345 1346 bzero(dropch, sizeof(dropch)); 1347 1348 /* 1349 * Replace chains in ip->cluster with chains from cluster and 1350 * adjust the focus if necessary. 1351 * 1352 * NOTE: nchain and/or ochain can be NULL due to gaps 1353 * in the cluster arrays. 1354 */ 1355 hammer2_spin_ex(&ip->cluster_spin); 1356 for (i = 0; cluster && i < cluster->nchains; ++i) { 1357 /* 1358 * Do not replace elements which are the same. Also handle 1359 * element count discrepancies. 1360 */ 1361 nchain = cluster->array[i].chain; 1362 if (i < ip->cluster.nchains) { 1363 ochain = ip->cluster.array[i].chain; 1364 if (ochain == nchain) 1365 continue; 1366 } else { 1367 ochain = NULL; 1368 } 1369 1370 /* 1371 * Make adjustments 1372 */ 1373 ip->cluster.array[i].chain = nchain; 1374 ip->cluster.array[i].flags &= ~HAMMER2_CITEM_INVALID; 1375 ip->cluster.array[i].flags |= cluster->array[i].flags & 1376 HAMMER2_CITEM_INVALID; 1377 if (nchain) 1378 hammer2_chain_ref(nchain); 1379 dropch[i] = ochain; 1380 } 1381 1382 /* 1383 * Release any left-over chains in ip->cluster. 1384 */ 1385 while (i < ip->cluster.nchains) { 1386 nchain = ip->cluster.array[i].chain; 1387 if (nchain) { 1388 ip->cluster.array[i].chain = NULL; 1389 ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID; 1390 } 1391 dropch[i] = nchain; 1392 ++i; 1393 } 1394 1395 /* 1396 * Fixup fields. Note that the inode-embedded cluster is never 1397 * directly locked. 1398 */ 1399 if (cluster) { 1400 ip->cluster.nchains = cluster->nchains; 1401 ip->cluster.focus = cluster->focus; 1402 ip->cluster.flags = cluster->flags & ~HAMMER2_CLUSTER_LOCKED; 1403 } else { 1404 ip->cluster.nchains = 0; 1405 ip->cluster.focus = NULL; 1406 ip->cluster.flags &= ~HAMMER2_CLUSTER_ZFLAGS; 1407 } 1408 1409 hammer2_spin_unex(&ip->cluster_spin); 1410 1411 /* 1412 * Cleanup outside of spinlock 1413 */ 1414 while (--i >= 0) { 1415 if (dropch[i]) 1416 hammer2_chain_drop(dropch[i]); 1417 } 1418 } 1419 1420 /* 1421 * Repoint a single element from the cluster to the ip. Used by the 1422 * synchronization threads to piecemeal update inodes. Does not change 1423 * focus and requires inode to be re-locked to clean-up flags (XXX). 1424 */ 1425 void 1426 hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster, 1427 int idx) 1428 { 1429 hammer2_chain_t *ochain; 1430 hammer2_chain_t *nchain; 1431 int i; 1432 1433 hammer2_spin_ex(&ip->cluster_spin); 1434 KKASSERT(idx < cluster->nchains); 1435 if (idx < ip->cluster.nchains) { 1436 ochain = ip->cluster.array[idx].chain; 1437 nchain = cluster->array[idx].chain; 1438 } else { 1439 ochain = NULL; 1440 nchain = cluster->array[idx].chain; 1441 for (i = ip->cluster.nchains; i <= idx; ++i) { 1442 bzero(&ip->cluster.array[i], 1443 sizeof(ip->cluster.array[i])); 1444 ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID; 1445 } 1446 ip->cluster.nchains = idx + 1; 1447 } 1448 if (ochain != nchain) { 1449 /* 1450 * Make adjustments. 1451 */ 1452 ip->cluster.array[idx].chain = nchain; 1453 ip->cluster.array[idx].flags &= ~HAMMER2_CITEM_INVALID; 1454 ip->cluster.array[idx].flags |= cluster->array[idx].flags & 1455 HAMMER2_CITEM_INVALID; 1456 } 1457 hammer2_spin_unex(&ip->cluster_spin); 1458 if (ochain != nchain) { 1459 if (nchain) 1460 hammer2_chain_ref(nchain); 1461 if (ochain) 1462 hammer2_chain_drop(ochain); 1463 } 1464 } 1465 1466 hammer2_key_t 1467 hammer2_inode_data_count(const hammer2_inode_t *ip) 1468 { 1469 hammer2_chain_t *chain; 1470 hammer2_key_t count = 0; 1471 int i; 1472 1473 for (i = 0; i < ip->cluster.nchains; ++i) { 1474 if ((chain = ip->cluster.array[i].chain) != NULL) { 1475 if (count < chain->bref.embed.stats.data_count) 1476 count = chain->bref.embed.stats.data_count; 1477 } 1478 } 1479 return count; 1480 } 1481 1482 hammer2_key_t 1483 hammer2_inode_inode_count(const hammer2_inode_t *ip) 1484 { 1485 hammer2_chain_t *chain; 1486 hammer2_key_t count = 0; 1487 int i; 1488 1489 for (i = 0; i < ip->cluster.nchains; ++i) { 1490 if ((chain = ip->cluster.array[i].chain) != NULL) { 1491 if (count < chain->bref.embed.stats.inode_count) 1492 count = chain->bref.embed.stats.inode_count; 1493 } 1494 } 1495 return count; 1496 } 1497 1498 /* 1499 * Called with a locked inode to finish unlinking an inode after xop_unlink 1500 * had been run. This function is responsible for decrementing nlinks. 1501 * 1502 * We don't bother decrementing nlinks if the file is not open and this was 1503 * the last link. 1504 * 1505 * If the inode is a hardlink target it's chain has not yet been deleted, 1506 * otherwise it's chain has been deleted. 1507 * 1508 * If isopen then any prior deletion was not permanent and the inode is 1509 * left intact with nlinks == 0; 1510 */ 1511 int 1512 hammer2_inode_unlink_finisher(hammer2_inode_t *ip, struct vnode **vprecyclep) 1513 { 1514 struct vnode *vp; 1515 1516 /* 1517 * Decrement nlinks. Catch a bad nlinks count here too (e.g. 0 or 1518 * negative), and just assume a transition to 0. 1519 */ 1520 if ((int64_t)ip->meta.nlinks <= 1) { 1521 atomic_set_int(&ip->flags, HAMMER2_INODE_ISUNLINKED); 1522 1523 /* 1524 * Scrap the vnode as quickly as possible. The vp association 1525 * stays intact while we hold the inode locked. However, vp 1526 * can be NULL here. 1527 */ 1528 vp = ip->vp; 1529 cpu_ccfence(); 1530 1531 /* 1532 * If no vp is associated there is no high-level state to 1533 * deal with and we can scrap the inode immediately. 1534 */ 1535 if (vp == NULL) { 1536 if ((ip->flags & HAMMER2_INODE_DELETING) == 0) { 1537 atomic_set_int(&ip->flags, 1538 HAMMER2_INODE_DELETING); 1539 hammer2_inode_delayed_sideq(ip); 1540 } 1541 return 0; 1542 } 1543 1544 /* 1545 * Because INODE_ISUNLINKED is set with the inode lock 1546 * held, the vnode cannot be ripped up from under us. 1547 * There may still be refs so knote anyone waiting for 1548 * a delete notification. 1549 * 1550 * The vnode is not necessarily ref'd due to the unlinking 1551 * itself, so we have to defer handling to the end of the 1552 * VOP, which will then call hammer2_inode_vprecycle(). 1553 */ 1554 if (vprecyclep) { 1555 vhold(vp); 1556 *vprecyclep = vp; 1557 } 1558 } 1559 1560 /* 1561 * Adjust nlinks and retain the inode on the media for now 1562 */ 1563 hammer2_inode_modify(ip); 1564 if ((int64_t)ip->meta.nlinks > 1) 1565 --ip->meta.nlinks; 1566 else 1567 ip->meta.nlinks = 0; 1568 1569 return 0; 1570 } 1571 1572 /* 1573 * Called at the end of a VOP that removes a file with a vnode that 1574 * we want to try to dispose of quickly due to a file deletion. If 1575 * we don't do this, the vnode can hang around with 0 refs for a very 1576 * long time and prevent reclamation of the underlying file and inode 1577 * (inode remains on-media with nlinks == 0 until the vnode is recycled 1578 * due to random system activity or a umount). 1579 */ 1580 void 1581 hammer2_inode_vprecycle(struct vnode *vp) 1582 { 1583 if (vget(vp, LK_EXCLUSIVE) == 0) { 1584 vfinalize(vp); 1585 hammer2_knote(vp, NOTE_DELETE); 1586 vdrop(vp); 1587 vput(vp); 1588 } else { 1589 vdrop(vp); 1590 } 1591 } 1592 1593 1594 /* 1595 * Mark an inode as being modified, meaning that the caller will modify 1596 * ip->meta. 1597 * 1598 * If a vnode is present we set the vnode dirty and the nominal filesystem 1599 * sync will also handle synchronizing the inode meta-data. Unless NOSIDEQ 1600 * we must ensure that the inode is on pmp->sideq. 1601 * 1602 * NOTE: We must always queue the inode to the sideq. This allows H2 to 1603 * shortcut vsyncscan() and flush inodes and their related vnodes 1604 * in a two stages. H2 still calls vfsync() for each vnode. 1605 * 1606 * NOTE: No mtid (modify_tid) is passed into this routine. The caller is 1607 * only modifying the in-memory inode. A modify_tid is synchronized 1608 * later when the inode gets flushed. 1609 * 1610 * NOTE: As an exception to the general rule, the inode MAY be locked 1611 * shared for this particular call. 1612 */ 1613 void 1614 hammer2_inode_modify(hammer2_inode_t *ip) 1615 { 1616 atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED); 1617 if (ip->vp) 1618 vsetisdirty(ip->vp); 1619 if (ip->pmp && (ip->flags & HAMMER2_INODE_NOSIDEQ) == 0) 1620 hammer2_inode_delayed_sideq(ip); 1621 } 1622 1623 /* 1624 * Synchronize the inode's frontend state with the chain state prior 1625 * to any explicit flush of the inode or any strategy write call. This 1626 * does not flush the inode's chain or its sub-topology to media (higher 1627 * level layers are responsible for doing that). 1628 * 1629 * Called with a locked inode inside a normal transaction. 1630 * 1631 * inode must be locked. 1632 */ 1633 int 1634 hammer2_inode_chain_sync(hammer2_inode_t *ip) 1635 { 1636 int error; 1637 1638 error = 0; 1639 if (ip->flags & (HAMMER2_INODE_RESIZED | HAMMER2_INODE_MODIFIED)) { 1640 hammer2_xop_fsync_t *xop; 1641 1642 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1643 xop->clear_directdata = 0; 1644 if (ip->flags & HAMMER2_INODE_RESIZED) { 1645 if ((ip->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) && 1646 ip->meta.size > HAMMER2_EMBEDDED_BYTES) { 1647 ip->meta.op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA; 1648 xop->clear_directdata = 1; 1649 } 1650 xop->osize = ip->osize; 1651 } else { 1652 xop->osize = ip->meta.size; /* safety */ 1653 } 1654 xop->ipflags = ip->flags; 1655 xop->meta = ip->meta; 1656 1657 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED | 1658 HAMMER2_INODE_MODIFIED); 1659 hammer2_xop_start(&xop->head, &hammer2_inode_chain_sync_desc); 1660 error = hammer2_xop_collect(&xop->head, 0); 1661 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1662 if (error == HAMMER2_ERROR_ENOENT) 1663 error = 0; 1664 if (error) { 1665 kprintf("hammer2: unable to fsync inode %p\n", ip); 1666 /* 1667 atomic_set_int(&ip->flags, 1668 xop->ipflags & (HAMMER2_INODE_RESIZED | 1669 HAMMER2_INODE_MODIFIED)); 1670 */ 1671 /* XXX return error somehow? */ 1672 } 1673 } 1674 return error; 1675 } 1676 1677 /* 1678 * When an inode is flagged INODE_CREATING its chains have not actually 1679 * been inserting into the on-media tree yet. 1680 */ 1681 int 1682 hammer2_inode_chain_ins(hammer2_inode_t *ip) 1683 { 1684 int error; 1685 1686 error = 0; 1687 if (ip->flags & HAMMER2_INODE_CREATING) { 1688 hammer2_xop_create_t *xop; 1689 1690 atomic_clear_int(&ip->flags, HAMMER2_INODE_CREATING); 1691 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1692 xop->lhc = ip->meta.inum; 1693 xop->flags = 0; 1694 hammer2_xop_start(&xop->head, &hammer2_inode_create_ins_desc); 1695 error = hammer2_xop_collect(&xop->head, 0); 1696 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1697 if (error == HAMMER2_ERROR_ENOENT) 1698 error = 0; 1699 if (error) { 1700 kprintf("hammer2: backend unable to " 1701 "insert inode %p %ld\n", ip, (long)ip->meta.inum); 1702 /* XXX return error somehow? */ 1703 } 1704 } 1705 return error; 1706 } 1707 1708 /* 1709 * When an inode is flagged INODE_DELETING it has been deleted (no directory 1710 * entry or open refs are left, though as an optimization H2 might leave 1711 * nlinks == 1 to avoid unnecessary block updates). The backend flush then 1712 * needs to actually remove it from the topology. 1713 * 1714 * NOTE: backend flush must still sync and flush the deleted inode to clean 1715 * out related chains. 1716 * 1717 * NOTE: We must clear not only INODE_DELETING, but also INODE_ISUNLINKED 1718 * to prevent the vnode reclaim code from trying to delete it twice. 1719 */ 1720 int 1721 hammer2_inode_chain_des(hammer2_inode_t *ip) 1722 { 1723 int error; 1724 1725 error = 0; 1726 if (ip->flags & HAMMER2_INODE_DELETING) { 1727 hammer2_xop_destroy_t *xop; 1728 1729 atomic_clear_int(&ip->flags, HAMMER2_INODE_DELETING | 1730 HAMMER2_INODE_ISUNLINKED); 1731 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING); 1732 hammer2_xop_start(&xop->head, &hammer2_inode_destroy_desc); 1733 error = hammer2_xop_collect(&xop->head, 0); 1734 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1735 1736 if (error == HAMMER2_ERROR_ENOENT) 1737 error = 0; 1738 if (error) { 1739 kprintf("hammer2: backend unable to " 1740 "delete inode %p %ld\n", ip, (long)ip->meta.inum); 1741 /* XXX return error somehow? */ 1742 } 1743 } 1744 return error; 1745 } 1746 1747 /* 1748 * Flushes the inode's chain and its sub-topology to media. Interlocks 1749 * HAMMER2_INODE_DIRTYDATA by clearing it prior to the flush. Any strategy 1750 * function creating or modifying a chain under this inode will re-set the 1751 * flag. 1752 * 1753 * inode must be locked. 1754 */ 1755 int 1756 hammer2_inode_chain_flush(hammer2_inode_t *ip, int flags) 1757 { 1758 hammer2_xop_fsync_t *xop; 1759 int error; 1760 1761 atomic_clear_int(&ip->flags, HAMMER2_INODE_DIRTYDATA); 1762 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING | flags); 1763 hammer2_xop_start(&xop->head, &hammer2_inode_flush_desc); 1764 error = hammer2_xop_collect(&xop->head, HAMMER2_XOP_COLLECT_WAITALL); 1765 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP); 1766 if (error == HAMMER2_ERROR_ENOENT) 1767 error = 0; 1768 1769 return error; 1770 } 1771