sys/vfs/hammer2/hammer2_inode.c

   1 /*
   2  * Copyright (c) 2011-2014 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
  42 #include "hammer2.h"
  43
  44 #define INODE_DEBUG     0
  45
  46 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
  47              hammer2_tid_t, meta.inum);
  48
  49 int
  50 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2)
  51 {
  52         if (ip1->meta.inum < ip2->meta.inum)
  53                 return(-1);
  54         if (ip1->meta.inum > ip2->meta.inum)
  55                 return(1);
  56         return(0);
  57 }
  58
  59 /*
  60  * HAMMER2 inode locks
  61  *
  62  * HAMMER2 offers shared and exclusive locks on inodes.  Pass a mask of
  63  * flags for options:
  64  *
  65  *      - pass HAMMER2_RESOLVE_SHARED if a shared lock is desired.  The
  66  *        inode locking function will automatically set the RDONLY flag.
  67  *
  68  *      - pass HAMMER2_RESOLVE_ALWAYS if you need the inode's meta-data.
  69  *        Most front-end inode locks do.
  70  *
  71  *      - pass HAMMER2_RESOLVE_NEVER if you do not want to require that
  72  *        the inode data be resolved.  This is used by the syncthr because
  73  *        it can run on an unresolved/out-of-sync cluster, and also by the
  74  *        vnode reclamation code to avoid unnecessary I/O (particularly when
  75  *        disposing of hundreds of thousands of cached vnodes).
  76  *
  77  * The inode locking function locks the inode itself, resolves any stale
  78  * chains in the inode's cluster, and allocates a fresh copy of the
  79  * cluster with 1 ref and all the underlying chains locked.
  80  *
  81  * ip->cluster will be stable while the inode is locked.
  82  *
  83  * NOTE: We don't combine the inode/chain lock because putting away an
  84  *       inode would otherwise confuse multiple lock holders of the inode.
  85  *
  86  * NOTE: In-memory inodes always point to hardlink targets (the actual file),
  87  *       and never point to a hardlink pointer.
  88  *
  89  * NOTE: If caller passes HAMMER2_RESOLVE_RDONLY the exclusive locking code
  90  *       will feel free to reduce the chain set in the cluster as an
  91  *       optimization.  It will still be validated against the quorum if
  92  *       appropriate, but the optimization might be able to reduce data
  93  *       accesses to one node.  This flag is automatically set if the inode
  94  *       is locked with HAMMER2_RESOLVE_SHARED.
  95  */
  96 void
  97 hammer2_inode_lock(hammer2_inode_t *ip, int how)
  98 {
  99         hammer2_inode_ref(ip);
 100
 101         /*
 102          * Inode structure mutex
 103          */
 104         if (how & HAMMER2_RESOLVE_SHARED) {
 105                 /*how |= HAMMER2_RESOLVE_RDONLY; not used */
 106                 hammer2_mtx_sh(&ip->lock);
 107         } else {
 108                 hammer2_mtx_ex(&ip->lock);
 109         }
 110 }
 111
 112 /*
 113  * Select a chain out of an inode's cluster and lock it.
 114  *
 115  * The inode does not have to be locked.
 116  */
 117 hammer2_chain_t *
 118 hammer2_inode_chain(hammer2_inode_t *ip, int clindex, int how)
 119 {
 120         hammer2_chain_t *chain;
 121
 122         hammer2_spin_sh(&ip->cluster_spin);
 123         if (clindex >= ip->cluster.nchains)
 124                 chain = NULL;
 125         else
 126                 chain = ip->cluster.array[clindex].chain;
 127         if (chain) {
 128                 hammer2_chain_ref(chain);
 129                 hammer2_spin_unsh(&ip->cluster_spin);
 130                 hammer2_chain_lock(chain, how);
 131         } else {
 132                 hammer2_spin_unsh(&ip->cluster_spin);
 133         }
 134         return chain;
 135 }
 136
 137 hammer2_chain_t *
 138 hammer2_inode_chain_and_parent(hammer2_inode_t *ip, int clindex,
 139                                hammer2_chain_t **parentp, int how)
 140 {
 141         hammer2_chain_t *chain;
 142         hammer2_chain_t *parent;
 143
 144         for (;;) {
 145                 hammer2_spin_sh(&ip->cluster_spin);
 146                 if (clindex >= ip->cluster.nchains)
 147                         chain = NULL;
 148                 else
 149                         chain = ip->cluster.array[clindex].chain;
 150                 if (chain) {
 151                         hammer2_chain_ref(chain);
 152                         hammer2_spin_unsh(&ip->cluster_spin);
 153                         hammer2_chain_lock(chain, how);
 154                 } else {
 155                         hammer2_spin_unsh(&ip->cluster_spin);
 156                 }
 157
 158                 /*
 159                  * Get parent, lock order must be (parent, chain).
 160                  */
 161                 parent = chain->parent;
 162                 hammer2_chain_ref(parent);
 163                 hammer2_chain_unlock(chain);
 164                 hammer2_chain_lock(parent, how);
 165                 hammer2_chain_lock(chain, how);
 166                 if (ip->cluster.array[clindex].chain == chain &&
 167                     chain->parent == parent) {
 168                         break;
 169                 }
 170
 171                 /*
 172                  * Retry
 173                  */
 174                 hammer2_chain_unlock(chain);
 175                 hammer2_chain_drop(chain);
 176                 hammer2_chain_unlock(parent);
 177                 hammer2_chain_drop(parent);
 178         }
 179         *parentp = parent;
 180
 181         return chain;
 182 }
 183
 184 void
 185 hammer2_inode_unlock(hammer2_inode_t *ip)
 186 {
 187         hammer2_mtx_unlock(&ip->lock);
 188         hammer2_inode_drop(ip);
 189 }
 190
 191 /*
 192  * Temporarily release a lock held shared or exclusive.  Caller must
 193  * hold the lock shared or exclusive on call and lock will be released
 194  * on return.
 195  *
 196  * Restore a lock that was temporarily released.
 197  */
 198 hammer2_mtx_state_t
 199 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
 200 {
 201         return hammer2_mtx_temp_release(&ip->lock);
 202 }
 203
 204 void
 205 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, hammer2_mtx_state_t ostate)
 206 {
 207         hammer2_mtx_temp_restore(&ip->lock, ostate);
 208 }
 209
 210 /*
 211  * Upgrade a shared inode lock to exclusive and return.  If the inode lock
 212  * is already held exclusively this is a NOP.
 213  *
 214  * The caller MUST hold the inode lock either shared or exclusive on call
 215  * and will own the lock exclusively on return.
 216  *
 217  * Returns non-zero if the lock was already exclusive prior to the upgrade.
 218  */
 219 int
 220 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
 221 {
 222         int wasexclusive;
 223
 224         if (mtx_islocked_ex(&ip->lock)) {
 225                 wasexclusive = 1;
 226         } else {
 227                 hammer2_mtx_unlock(&ip->lock);
 228                 hammer2_mtx_ex(&ip->lock);
 229                 wasexclusive = 0;
 230         }
 231         return wasexclusive;
 232 }
 233
 234 /*
 235  * Downgrade an inode lock from exclusive to shared only if the inode
 236  * lock was previously shared.  If the inode lock was previously exclusive,
 237  * this is a NOP.
 238  */
 239 void
 240 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, int wasexclusive)
 241 {
 242         if (wasexclusive == 0)
 243                 mtx_downgrade(&ip->lock);
 244 }
 245
 246 /*
 247  * Lookup an inode by inode number
 248  */
 249 hammer2_inode_t *
 250 hammer2_inode_lookup(hammer2_pfs_t *pmp, hammer2_tid_t inum)
 251 {
 252         hammer2_inode_t *ip;
 253
 254         KKASSERT(pmp);
 255         if (pmp->spmp_hmp) {
 256                 ip = NULL;
 257         } else {
 258                 hammer2_spin_ex(&pmp->inum_spin);
 259                 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
 260                 if (ip)
 261                         hammer2_inode_ref(ip);
 262                 hammer2_spin_unex(&pmp->inum_spin);
 263         }
 264         return(ip);
 265 }
 266
 267 /*
 268  * Adding a ref to an inode is only legal if the inode already has at least
 269  * one ref.
 270  *
 271  * (can be called with spinlock held)
 272  */
 273 void
 274 hammer2_inode_ref(hammer2_inode_t *ip)
 275 {
 276         atomic_add_int(&ip->refs, 1);
 277 }
 278
 279 /*
 280  * Drop an inode reference, freeing the inode when the last reference goes
 281  * away.
 282  */
 283 void
 284 hammer2_inode_drop(hammer2_inode_t *ip)
 285 {
 286         hammer2_pfs_t *pmp;
 287         hammer2_inode_t *pip;
 288         u_int refs;
 289
 290         while (ip) {
 291                 refs = ip->refs;
 292                 cpu_ccfence();
 293                 if (refs == 1) {
 294                         /*
 295                          * Transition to zero, must interlock with
 296                          * the inode inumber lookup tree (if applicable).
 297                          * It should not be possible for anyone to race
 298                          * the transition to 0.
 299                          */
 300                         pmp = ip->pmp;
 301                         KKASSERT(pmp);
 302                         hammer2_spin_ex(&pmp->inum_spin);
 303
 304                         if (atomic_cmpset_int(&ip->refs, 1, 0)) {
 305                                 KKASSERT(hammer2_mtx_refs(&ip->lock) == 0);
 306                                 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
 307                                         atomic_clear_int(&ip->flags,
 308                                                      HAMMER2_INODE_ONRBTREE);
 309                                         RB_REMOVE(hammer2_inode_tree,
 310                                                   &pmp->inum_tree, ip);
 311                                 }
 312                                 hammer2_spin_unex(&pmp->inum_spin);
 313
 314                                 pip = ip->pip;
 315                                 ip->pip = NULL;
 316                                 ip->pmp = NULL;
 317
 318                                 /*
 319                                  * Cleaning out ip->cluster isn't entirely
 320                                  * trivial.
 321                                  */
 322                                 hammer2_inode_repoint(ip, NULL, NULL);
 323
 324                                 /*
 325                                  * We have to drop pip (if non-NULL) to
 326                                  * dispose of our implied reference from
 327                                  * ip->pip.  We can simply loop on it.
 328                                  */
 329                                 kfree(ip, pmp->minode);
 330                                 atomic_add_long(&pmp->inmem_inodes, -1);
 331                                 ip = pip;
 332                                 /* continue with pip (can be NULL) */
 333                         } else {
 334                                 hammer2_spin_unex(&ip->pmp->inum_spin);
 335                         }
 336                 } else {
 337                         /*
 338                          * Non zero transition
 339                          */
 340                         if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
 341                                 break;
 342                 }
 343         }
 344 }
 345
 346 /*
 347  * Get the vnode associated with the given inode, allocating the vnode if
 348  * necessary.  The vnode will be returned exclusively locked.
 349  *
 350  * The caller must lock the inode (shared or exclusive).
 351  *
 352  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
 353  * races.
 354  */
 355 struct vnode *
 356 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
 357 {
 358         hammer2_pfs_t *pmp;
 359         struct vnode *vp;
 360
 361         pmp = ip->pmp;
 362         KKASSERT(pmp != NULL);
 363         *errorp = 0;
 364
 365         for (;;) {
 366                 /*
 367                  * Attempt to reuse an existing vnode assignment.  It is
 368                  * possible to race a reclaim so the vget() may fail.  The
 369                  * inode must be unlocked during the vget() to avoid a
 370                  * deadlock against a reclaim.
 371                  */
 372                 int wasexclusive;
 373
 374                 vp = ip->vp;
 375                 if (vp) {
 376                         /*
 377                          * Inode must be unlocked during the vget() to avoid
 378                          * possible deadlocks, but leave the ip ref intact.
 379                          *
 380                          * vnode is held to prevent destruction during the
 381                          * vget().  The vget() can still fail if we lost
 382                          * a reclaim race on the vnode.
 383                          */
 384                         hammer2_mtx_state_t ostate;
 385
 386                         vhold(vp);
 387                         ostate = hammer2_inode_lock_temp_release(ip);
 388                         if (vget(vp, LK_EXCLUSIVE)) {
 389                                 vdrop(vp);
 390                                 hammer2_inode_lock_temp_restore(ip, ostate);
 391                                 continue;
 392                         }
 393                         hammer2_inode_lock_temp_restore(ip, ostate);
 394                         vdrop(vp);
 395                         /* vp still locked and ref from vget */
 396                         if (ip->vp != vp) {
 397                                 kprintf("hammer2: igetv race %p/%p\n",
 398                                         ip->vp, vp);
 399                                 vput(vp);
 400                                 continue;
 401                         }
 402                         *errorp = 0;
 403                         break;
 404                 }
 405
 406                 /*
 407                  * No vnode exists, allocate a new vnode.  Beware of
 408                  * allocation races.  This function will return an
 409                  * exclusively locked and referenced vnode.
 410                  */
 411                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
 412                 if (*errorp) {
 413                         kprintf("hammer2: igetv getnewvnode failed %d\n",
 414                                 *errorp);
 415                         vp = NULL;
 416                         break;
 417                 }
 418
 419                 /*
 420                  * Lock the inode and check for an allocation race.
 421                  */
 422                 wasexclusive = hammer2_inode_lock_upgrade(ip);
 423                 if (ip->vp != NULL) {
 424                         vp->v_type = VBAD;
 425                         vx_put(vp);
 426                         hammer2_inode_lock_downgrade(ip, wasexclusive);
 427                         continue;
 428                 }
 429
 430                 switch (ip->meta.type) {
 431                 case HAMMER2_OBJTYPE_DIRECTORY:
 432                         vp->v_type = VDIR;
 433                         break;
 434                 case HAMMER2_OBJTYPE_REGFILE:
 435                         vp->v_type = VREG;
 436                         vinitvmio(vp, ip->meta.size,
 437                                   HAMMER2_LBUFSIZE,
 438                                   (int)ip->meta.size & HAMMER2_LBUFMASK);
 439                         break;
 440                 case HAMMER2_OBJTYPE_SOFTLINK:
 441                         /*
 442                          * XXX for now we are using the generic file_read
 443                          * and file_write code so we need a buffer cache
 444                          * association.
 445                          */
 446                         vp->v_type = VLNK;
 447                         vinitvmio(vp, ip->meta.size,
 448                                   HAMMER2_LBUFSIZE,
 449                                   (int)ip->meta.size & HAMMER2_LBUFMASK);
 450                         break;
 451                 case HAMMER2_OBJTYPE_CDEV:
 452                         vp->v_type = VCHR;
 453                         /* fall through */
 454                 case HAMMER2_OBJTYPE_BDEV:
 455                         vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
 456                         if (ip->meta.type != HAMMER2_OBJTYPE_CDEV)
 457                                 vp->v_type = VBLK;
 458                         addaliasu(vp,
 459                                   ip->meta.rmajor,
 460                                   ip->meta.rminor);
 461                         break;
 462                 case HAMMER2_OBJTYPE_FIFO:
 463                         vp->v_type = VFIFO;
 464                         vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
 465                         break;
 466                 default:
 467                         panic("hammer2: unhandled objtype %d",
 468                               ip->meta.type);
 469                         break;
 470                 }
 471
 472                 if (ip == pmp->iroot)
 473                         vsetflags(vp, VROOT);
 474
 475                 vp->v_data = ip;
 476                 ip->vp = vp;
 477                 hammer2_inode_ref(ip);          /* vp association */
 478                 hammer2_inode_lock_downgrade(ip, wasexclusive);
 479                 break;
 480         }
 481
 482         /*
 483          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
 484          */
 485         if (hammer2_debug & 0x0002) {
 486                 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
 487                         vp, vp->v_refcnt, vp->v_auxrefs);
 488         }
 489         return (vp);
 490 }
 491
 492 /*
 493  * Returns the inode associated with the passed-in cluster, creating the
 494  * inode if necessary and synchronizing it to the passed-in cluster otherwise.
 495  * When synchronizing, if idx >= 0, only cluster index (idx) is synchronized.
 496  * Otherwise the whole cluster is synchronized.
 497  *
 498  * The passed-in cluster must be locked and will remain locked on return.
 499  * The returned inode will be locked and the caller may dispose of both
 500  * via hammer2_inode_unlock() + hammer2_inode_drop().  However, if the caller
 501  * needs to resolve a hardlink it must ref/unlock/relock/drop the inode.
 502  *
 503  * The hammer2_inode structure regulates the interface between the high level
 504  * kernel VNOPS API and the filesystem backend (the chains).
 505  *
 506  * On return the inode is locked with the supplied cluster.
 507  */
 508 hammer2_inode_t *
 509 hammer2_inode_get(hammer2_pfs_t *pmp, hammer2_inode_t *dip,
 510                   hammer2_cluster_t *cluster, int idx)
 511 {
 512         hammer2_inode_t *nip;
 513         const hammer2_inode_data_t *iptmp;
 514         const hammer2_inode_data_t *nipdata;
 515
 516         KKASSERT(cluster == NULL ||
 517                  hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE);
 518         KKASSERT(pmp);
 519
 520         /*
 521          * Interlocked lookup/ref of the inode.  This code is only needed
 522          * when looking up inodes with nlinks != 0 (TODO: optimize out
 523          * otherwise and test for duplicates).
 524          *
 525          * Cluster can be NULL during the initial pfs allocation.
 526          */
 527 again:
 528         while (cluster) {
 529                 iptmp = &hammer2_cluster_rdata(cluster)->ipdata;
 530                 nip = hammer2_inode_lookup(pmp, iptmp->meta.inum);
 531                 if (nip == NULL)
 532                         break;
 533
 534                 hammer2_mtx_ex(&nip->lock);
 535
 536                 /*
 537                  * Handle SMP race (not applicable to the super-root spmp
 538                  * which can't index inodes due to duplicative inode numbers).
 539                  */
 540                 if (pmp->spmp_hmp == NULL &&
 541                     (nip->flags & HAMMER2_INODE_ONRBTREE) == 0) {
 542                         hammer2_mtx_unlock(&nip->lock);
 543                         hammer2_inode_drop(nip);
 544                         continue;
 545                 }
 546                 if (idx >= 0)
 547                         hammer2_inode_repoint_one(nip, cluster, idx);
 548                 else
 549                         hammer2_inode_repoint(nip, NULL, cluster);
 550
 551                 return nip;
 552         }
 553
 554         /*
 555          * We couldn't find the inode number, create a new inode.
 556          */
 557         nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
 558         spin_init(&nip->cluster_spin, "h2clspin");
 559         atomic_add_long(&pmp->inmem_inodes, 1);
 560         hammer2_pfs_memory_inc(pmp);
 561         hammer2_pfs_memory_wakeup(pmp);
 562         if (pmp->spmp_hmp)
 563                 nip->flags = HAMMER2_INODE_SROOT;
 564
 565         /*
 566          * Initialize nip's cluster.  A cluster is provided for normal
 567          * inodes but typically not for the super-root or PFS inodes.
 568          */
 569         nip->cluster.refs = 1;
 570         nip->cluster.pmp = pmp;
 571         nip->cluster.flags |= HAMMER2_CLUSTER_INODE;
 572         if (cluster) {
 573                 nipdata = &hammer2_cluster_rdata(cluster)->ipdata;
 574                 nip->meta = nipdata->meta;
 575                 atomic_set_int(&nip->flags, HAMMER2_INODE_METAGOOD);
 576                 hammer2_inode_repoint(nip, NULL, cluster);
 577         } else {
 578                 nip->meta.inum = 1;             /* PFS inum is always 1 XXX */
 579                 /* mtime will be updated when a cluster is available */
 580                 atomic_set_int(&nip->flags, HAMMER2_INODE_METAGOOD);/*XXX*/
 581         }
 582
 583         nip->pip = dip;                         /* can be NULL */
 584         if (dip)
 585                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
 586
 587         nip->pmp = pmp;
 588
 589         /*
 590          * ref and lock on nip gives it state compatible to after a
 591          * hammer2_inode_lock() call.
 592          */
 593         nip->refs = 1;
 594         hammer2_mtx_init(&nip->lock, "h2inode");
 595         hammer2_mtx_ex(&nip->lock);
 596         /* combination of thread lock and chain lock == inode lock */
 597
 598         /*
 599          * Attempt to add the inode.  If it fails we raced another inode
 600          * get.  Undo all the work and try again.
 601          */
 602         if (pmp->spmp_hmp == NULL) {
 603                 hammer2_spin_ex(&pmp->inum_spin);
 604                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
 605                         hammer2_spin_unex(&pmp->inum_spin);
 606                         hammer2_mtx_unlock(&nip->lock);
 607                         hammer2_inode_drop(nip);
 608                         goto again;
 609                 }
 610                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
 611                 hammer2_spin_unex(&pmp->inum_spin);
 612         }
 613
 614         return (nip);
 615 }
 616
 617 /*
 618  * Create a new inode in the specified directory using the vattr to
 619  * figure out the type of inode.
 620  *
 621  * If no error occurs the new inode with its cluster locked is returned in
 622  * *nipp, otherwise an error is returned and *nipp is set to NULL.
 623  *
 624  * If vap and/or cred are NULL the related fields are not set and the
 625  * inode type defaults to a directory.  This is used when creating PFSs
 626  * under the super-root, so the inode number is set to 1 in this case.
 627  *
 628  * dip is not locked on entry.
 629  *
 630  * NOTE: When used to create a snapshot, the inode is temporarily associated
 631  *       with the super-root spmp. XXX should pass new pmp for snapshot.
 632  */
 633 hammer2_inode_t *
 634 hammer2_inode_create(hammer2_inode_t *dip,
 635                      struct vattr *vap, struct ucred *cred,
 636                      const uint8_t *name, size_t name_len, hammer2_key_t lhc,
 637                      hammer2_key_t inum, uint8_t type, uint8_t target_type,
 638                      int flags, int *errorp)
 639 {
 640         hammer2_xop_create_t *xop;
 641         hammer2_inode_t *nip;
 642         int error;
 643         uid_t xuid;
 644         uuid_t dip_uid;
 645         uuid_t dip_gid;
 646         uint32_t dip_mode;
 647         uint8_t dip_comp_algo;
 648         uint8_t dip_check_algo;
 649
 650         if (name)
 651                 lhc = hammer2_dirhash(name, name_len);
 652         *errorp = 0;
 653         nip = NULL;
 654
 655         /*
 656          * Locate the inode or indirect block to create the new
 657          * entry in.  At the same time check for key collisions
 658          * and iterate until we don't get one.
 659          *
 660          * NOTE: hidden inodes do not have iterators.
 661          *
 662          * Lock the directory exclusively for now to guarantee that
 663          * we can find an unused lhc for the name.  Due to collisions,
 664          * two different creates can end up with the same lhc so we
 665          * cannot depend on the OS to prevent the collision.
 666          */
 667         hammer2_inode_lock(dip, 0);
 668
 669         dip_uid = dip->meta.uid;
 670         dip_gid = dip->meta.gid;
 671         dip_mode = dip->meta.mode;
 672         dip_comp_algo = dip->meta.comp_algo;
 673         dip_check_algo = dip->meta.check_algo;
 674
 675         /*
 676          * If name specified, locate an unused key in the collision space.
 677          * Otherwise use the passed-in lhc directly.
 678          */
 679         if (name) {
 680                 hammer2_xop_scanlhc_t *sxop;
 681                 hammer2_key_t lhcbase;
 682
 683                 lhcbase = lhc;
 684                 sxop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING);
 685                 sxop->lhc = lhc;
 686                 hammer2_xop_start(&sxop->head, hammer2_xop_scanlhc);
 687                 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) {
 688                         if (lhc != sxop->head.cluster.focus->bref.key)
 689                                 break;
 690                         ++lhc;
 691                 }
 692                 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP);
 693
 694                 if (error) {
 695                         if (error != ENOENT)
 696                                 goto done2;
 697                         ++lhc;
 698                         error = 0;
 699                 }
 700                 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) {
 701                         error = ENOSPC;
 702                         goto done2;
 703                 }
 704         }
 705
 706         /*
 707          * Create the inode with the lhc as the key.
 708          */
 709         xop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING);
 710         xop->lhc = lhc;
 711         xop->flags = flags;
 712         bzero(&xop->meta, sizeof(xop->meta));
 713
 714         if (vap) {
 715                 xop->meta.type = hammer2_get_obj_type(vap->va_type);
 716
 717                 switch (xop->meta.type) {
 718                 case HAMMER2_OBJTYPE_CDEV:
 719                 case HAMMER2_OBJTYPE_BDEV:
 720                         xop->meta.rmajor = vap->va_rmajor;
 721                         xop->meta.rminor = vap->va_rminor;
 722                         break;
 723                 default:
 724                         break;
 725                 }
 726                 type = xop->meta.type;
 727         } else {
 728                 xop->meta.type = type;
 729                 xop->meta.target_type = target_type;
 730         }
 731         xop->meta.inum = inum;
 732
 733         /* Inherit parent's inode compression mode. */
 734         xop->meta.comp_algo = dip_comp_algo;
 735         xop->meta.check_algo = dip_check_algo;
 736         xop->meta.version = HAMMER2_INODE_VERSION_ONE;
 737         hammer2_update_time(&xop->meta.ctime);
 738         xop->meta.mtime = xop->meta.ctime;
 739         if (vap)
 740                 xop->meta.mode = vap->va_mode;
 741         xop->meta.nlinks = 1;
 742         if (vap) {
 743                 if (dip && dip->pmp) {
 744                         xuid = hammer2_to_unix_xid(&dip_uid);
 745                         xuid = vop_helper_create_uid(dip->pmp->mp,
 746                                                      dip_mode,
 747                                                      xuid,
 748                                                      cred,
 749                                                      &vap->va_mode);
 750                 } else {
 751                         /* super-root has no dip and/or pmp */
 752                         xuid = 0;
 753                 }
 754                 if (vap->va_vaflags & VA_UID_UUID_VALID)
 755                         xop->meta.uid = vap->va_uid_uuid;
 756                 else if (vap->va_uid != (uid_t)VNOVAL)
 757                         hammer2_guid_to_uuid(&xop->meta.uid, vap->va_uid);
 758                 else
 759                         hammer2_guid_to_uuid(&xop->meta.uid, xuid);
 760
 761                 if (vap->va_vaflags & VA_GID_UUID_VALID)
 762                         xop->meta.gid = vap->va_gid_uuid;
 763                 else if (vap->va_gid != (gid_t)VNOVAL)
 764                         hammer2_guid_to_uuid(&xop->meta.gid, vap->va_gid);
 765                 else if (dip)
 766                         xop->meta.gid = dip_gid;
 767         }
 768
 769         /*
 770          * Regular files and softlinks allow a small amount of data to be
 771          * directly embedded in the inode.  This flag will be cleared if
 772          * the size is extended past the embedded limit.
 773          */
 774         if (xop->meta.type == HAMMER2_OBJTYPE_REGFILE ||
 775             xop->meta.type == HAMMER2_OBJTYPE_SOFTLINK ||
 776             xop->meta.type == HAMMER2_OBJTYPE_HARDLINK) {
 777                 xop->meta.op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
 778         }
 779         if (name)
 780                 hammer2_xop_setname(&xop->head, name, name_len);
 781         xop->meta.name_len = name_len;
 782         xop->meta.name_key = lhc;
 783         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 784
 785         hammer2_xop_start(&xop->head, hammer2_inode_xop_create);
 786
 787         error = hammer2_xop_collect(&xop->head, 0);
 788 #if INODE_DEBUG
 789         kprintf("CREATE INODE %*.*s\n",
 790                 (int)name_len, (int)name_len, name);
 791 #endif
 792
 793         if (error) {
 794                 *errorp = error;
 795                 goto done;
 796         }
 797
 798         /*
 799          * Set up the new inode if not a hardlink pointer.
 800          *
 801          * NOTE: *_get() integrates chain's lock into the inode lock.
 802          *
 803          * NOTE: Only one new inode can currently be created per
 804          *       transaction.  If the need arises we can adjust
 805          *       hammer2_trans_init() to allow more.
 806          *
 807          * NOTE: nipdata will have chain's blockset data.
 808          */
 809         if (type != HAMMER2_OBJTYPE_HARDLINK) {
 810                 nip = hammer2_inode_get(dip->pmp, dip, &xop->head.cluster, -1);
 811                 nip->comp_heuristic = 0;
 812         } else {
 813                 nip = NULL;
 814         }
 815
 816 done:
 817         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 818 done2:
 819         hammer2_inode_unlock(dip);
 820
 821         return (nip);
 822 }
 823
 824 /*
 825  * Connect the disconnected inode (ip) to the directory (dip) with the
 826  * specified (name, name_len).  If name is NULL, (lhc) will be used as
 827  * the directory key and the inode's embedded name will not be modified
 828  * for future recovery purposes.
 829  *
 830  * dip and ip must both be locked exclusively (dip in particular to avoid
 831  * lhc collisions).
 832  */
 833 int
 834 hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *ip,
 835                       const char *name, size_t name_len,
 836                       hammer2_key_t lhc)
 837 {
 838         hammer2_xop_scanlhc_t *sxop;
 839         hammer2_xop_connect_t *xop;
 840         hammer2_inode_t *opip;
 841         hammer2_key_t lhcbase;
 842         int error;
 843
 844         /*
 845          * Calculate the lhc and resolve the collision space.
 846          */
 847         if (name) {
 848                 lhc = lhcbase = hammer2_dirhash(name, name_len);
 849                 sxop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING);
 850                 sxop->lhc = lhc;
 851                 hammer2_xop_start(&sxop->head, hammer2_xop_scanlhc);
 852                 while ((error = hammer2_xop_collect(&sxop->head, 0)) == 0) {
 853                         if (lhc != sxop->head.cluster.focus->bref.key)
 854                                 break;
 855                         ++lhc;
 856                 }
 857                 hammer2_xop_retire(&sxop->head, HAMMER2_XOPMASK_VOP);
 858
 859                 if (error) {
 860                         if (error != ENOENT)
 861                                 goto done;
 862                         ++lhc;
 863                         error = 0;
 864                 }
 865                 if ((lhcbase ^ lhc) & ~HAMMER2_DIRHASH_LOMASK) {
 866                         error = ENOSPC;
 867                         goto done;
 868                 }
 869         } else {
 870                 error = 0;
 871         }
 872
 873         /*
 874          * Formally reconnect the in-memory structure.  ip must
 875          * be locked exclusively to safely change ip->pip.
 876          */
 877         if (ip->pip != dip) {
 878                 hammer2_inode_ref(dip);
 879                 opip = ip->pip;
 880                 ip->pip = dip;
 881                 if (opip)
 882                         hammer2_inode_drop(opip);
 883         }
 884
 885         /*
 886          * Connect her up
 887          */
 888         xop = hammer2_xop_alloc(dip, HAMMER2_XOP_MODIFYING);
 889         if (name)
 890                 hammer2_xop_setname(&xop->head, name, name_len);
 891         hammer2_xop_setip2(&xop->head, ip);
 892         xop->lhc = lhc;
 893         hammer2_xop_start(&xop->head, hammer2_inode_xop_connect);
 894         error = hammer2_xop_collect(&xop->head, 0);
 895         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
 896
 897         /*
 898          * On success make the same adjustments to ip->meta or the
 899          * next flush may blow up the chain.
 900          */
 901         if (error == 0) {
 902                 hammer2_inode_modify(ip);
 903                 ip->meta.name_key = lhc;
 904                 if (name)
 905                         ip->meta.name_len = name_len;
 906         }
 907 done:
 908         return error;
 909 }
 910
 911 /*
 912  * Repoint ip->cluster's chains to cluster's chains and fixup the default
 913  * focus.  All items, valid or invalid, are repointed.  hammer2_xop_start()
 914  * filters out invalid or non-matching elements.
 915  *
 916  * Caller must hold the inode and cluster exclusive locked, if not NULL,
 917  * must also be locked.
 918  *
 919  * Cluster may be NULL to clean out any chains in ip->cluster.
 920  */
 921 void
 922 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 923                       hammer2_cluster_t *cluster)
 924 {
 925         hammer2_chain_t *dropch[HAMMER2_MAXCLUSTER];
 926         hammer2_chain_t *ochain;
 927         hammer2_chain_t *nchain;
 928         hammer2_inode_t *opip;
 929         int i;
 930
 931         bzero(dropch, sizeof(dropch));
 932
 933         /*
 934          * Replace chains in ip->cluster with chains from cluster and
 935          * adjust the focus if necessary.
 936          *
 937          * NOTE: nchain and/or ochain can be NULL due to gaps
 938          *       in the cluster arrays.
 939          */
 940         hammer2_spin_ex(&ip->cluster_spin);
 941         for (i = 0; cluster && i < cluster->nchains; ++i) {
 942                 /*
 943                  * Do not replace elements which are the same.  Also handle
 944                  * element count discrepancies.
 945                  */
 946                 nchain = cluster->array[i].chain;
 947                 if (i < ip->cluster.nchains) {
 948                         ochain = ip->cluster.array[i].chain;
 949                         if (ochain == nchain)
 950                                 continue;
 951                 } else {
 952                         ochain = NULL;
 953                 }
 954
 955                 /*
 956                  * Make adjustments
 957                  */
 958                 ip->cluster.array[i].chain = nchain;
 959                 ip->cluster.array[i].flags &= ~HAMMER2_CITEM_INVALID;
 960                 ip->cluster.array[i].flags |= cluster->array[i].flags &
 961                                               HAMMER2_CITEM_INVALID;
 962                 if (nchain)
 963                         hammer2_chain_ref(nchain);
 964                 dropch[i] = ochain;
 965         }
 966
 967         /*
 968          * Release any left-over chains in ip->cluster.
 969          */
 970         while (i < ip->cluster.nchains) {
 971                 nchain = ip->cluster.array[i].chain;
 972                 if (nchain) {
 973                         ip->cluster.array[i].chain = NULL;
 974                         ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
 975                 }
 976                 dropch[i] = nchain;
 977                 ++i;
 978         }
 979
 980         /*
 981          * Fixup fields.  Note that the inode-embedded cluster is never
 982          * directly locked.
 983          */
 984         if (cluster) {
 985                 ip->cluster.nchains = cluster->nchains;
 986                 ip->cluster.focus = cluster->focus;
 987                 ip->cluster.flags = cluster->flags & ~HAMMER2_CLUSTER_LOCKED;
 988         } else {
 989                 ip->cluster.nchains = 0;
 990                 ip->cluster.focus = NULL;
 991                 ip->cluster.flags &= ~HAMMER2_CLUSTER_ZFLAGS;
 992         }
 993
 994         /*
 995          * Repoint ip->pip if requested (non-NULL pip).
 996          */
 997         if (pip && ip->pip != pip) {
 998                 opip = ip->pip;
 999                 hammer2_inode_ref(pip);
1000                 ip->pip = pip;
1001         } else {
1002                 opip = NULL;
1003         }
1004         hammer2_spin_unex(&ip->cluster_spin);
1005
1006         /*
1007          * Cleanup outside of spinlock
1008          */
1009         while (--i >= 0) {
1010                 if (dropch[i])
1011                         hammer2_chain_drop(dropch[i]);
1012         }
1013         if (opip)
1014                 hammer2_inode_drop(opip);
1015 }
1016
1017 /*
1018  * Repoint a single element from the cluster to the ip.  Used by the
1019  * synchronization threads to piecemeal update inodes.  Does not change
1020  * focus and requires inode to be re-locked to clean-up flags (XXX).
1021  */
1022 void
1023 hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
1024                           int idx)
1025 {
1026         hammer2_chain_t *ochain;
1027         hammer2_chain_t *nchain;
1028         int i;
1029
1030         hammer2_spin_ex(&ip->cluster_spin);
1031         KKASSERT(idx < cluster->nchains);
1032         if (idx < ip->cluster.nchains) {
1033                 ochain = ip->cluster.array[idx].chain;
1034                 nchain = cluster->array[idx].chain;
1035         } else {
1036                 ochain = NULL;
1037                 nchain = cluster->array[idx].chain;
1038                 ip->cluster.nchains = idx + 1;
1039                 for (i = ip->cluster.nchains; i <= idx; ++i) {
1040                         bzero(&ip->cluster.array[i],
1041                               sizeof(ip->cluster.array[i]));
1042                         ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
1043                 }
1044         }
1045         if (ochain != nchain) {
1046                 /*
1047                  * Make adjustments.
1048                  */
1049                 ip->cluster.array[idx].chain = nchain;
1050                 ip->cluster.array[idx].flags &= ~HAMMER2_CITEM_INVALID;
1051                 ip->cluster.array[idx].flags |= cluster->array[idx].flags &
1052                                                 HAMMER2_CITEM_INVALID;
1053         }
1054         hammer2_spin_unex(&ip->cluster_spin);
1055         if (ochain != nchain) {
1056                 if (nchain)
1057                         hammer2_chain_ref(nchain);
1058                 if (ochain)
1059                         hammer2_chain_drop(ochain);
1060         }
1061 }
1062
1063 /*
1064  * Called with a locked inode to finish unlinking an inode after xop_unlink
1065  * had been run.  This function is responsible for decrementing nlinks and
1066  * moving deleted inodes to the hidden directory if they are still open.
1067  *
1068  * We don't bother decrementing nlinks if the file is not open and this was
1069  * the last link.
1070  *
1071  * If the inode is a hardlink target it's chain has not yet been deleted,
1072  * otherwise it's chain has been deleted.
1073  *
1074  * If isopen then any prior deletion was not permanent and the inode must
1075  * be moved to the hidden directory.
1076  */
1077 int
1078 hammer2_inode_unlink_finisher(hammer2_inode_t *ip, int isopen)
1079 {
1080         hammer2_pfs_t *pmp;
1081         int error;
1082
1083         pmp = ip->pmp;
1084
1085         /*
1086          * Decrement nlinks.  If this is the last link and the file is
1087          * not open, the chain has already been removed and we don't bother
1088          * dirtying the inode.
1089          */
1090         if (ip->meta.nlinks == 1) {
1091                 atomic_set_int(&ip->flags, HAMMER2_INODE_ISUNLINKED);
1092                 if (isopen == 0)
1093                         return 0;
1094         }
1095
1096         hammer2_inode_modify(ip);
1097         --ip->meta.nlinks;
1098         if ((int64_t)ip->meta.nlinks < 0)
1099                 ip->meta.nlinks = 0;    /* safety */
1100
1101         /*
1102          * If nlinks is not zero we are done.  However, this should only be
1103          * possible with a hardlink target.  If the inode is an embedded
1104          * hardlink nlinks should have dropped to zero, warn and proceed
1105          * with the next step.
1106          */
1107         if (ip->meta.nlinks) {
1108                 if ((ip->meta.name_key & HAMMER2_DIRHASH_VISIBLE) == 0)
1109                         return 0;
1110                 kprintf("hammer2_inode_unlink: nlinks was not 0 (%jd)\n",
1111                         (intmax_t)ip->meta.nlinks);
1112                 return 0;
1113         }
1114
1115         /*
1116          * nlinks is now zero, the inode should have already been deleted.
1117          * If the file is open it was deleted non-permanently and must be
1118          * moved to the hidden directory.
1119          *
1120          * When moving to the hidden directory we force the name_key to the
1121          * inode number to avoid collisions.
1122          */
1123         if (isopen) {
1124                 hammer2_inode_lock(pmp->ihidden, 0);
1125                 error = hammer2_inode_connect(pmp->ihidden, ip,
1126                                               NULL, 0, ip->meta.inum);
1127                 hammer2_inode_unlock(pmp->ihidden);
1128         } else {
1129                 error = 0;
1130         }
1131         return error;
1132 }
1133
1134 /*
1135  * This is called from the mount code to initialize pmp->ihidden
1136  */
1137 void
1138 hammer2_inode_install_hidden(hammer2_pfs_t *pmp)
1139 {
1140         int error;
1141
1142         if (pmp->ihidden)
1143                 return;
1144
1145         hammer2_trans_init(pmp, 0);
1146         hammer2_inode_lock(pmp->iroot, 0);
1147
1148         /*
1149          * Find the hidden directory
1150          */
1151         {
1152                 hammer2_xop_lookup_t *xop;
1153
1154                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1155                 xop->lhc = HAMMER2_INODE_HIDDENDIR;
1156                 hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1157                 error = hammer2_xop_collect(&xop->head, 0);
1158
1159                 if (error == 0) {
1160                         /*
1161                          * Found the hidden directory
1162                          */
1163                         kprintf("PFS FOUND HIDDEN DIR\n");
1164                         pmp->ihidden = hammer2_inode_get(pmp, pmp->iroot,
1165                                                          &xop->head.cluster,
1166                                                          -1);
1167                         hammer2_inode_ref(pmp->ihidden);
1168                         hammer2_inode_unlock(pmp->ihidden);
1169                 }
1170                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1171         }
1172
1173         /*
1174          * Create the hidden directory if it could not be found.
1175          */
1176         if (error == ENOENT) {
1177                 kprintf("PFS CREATE HIDDEN DIR\n");
1178
1179                 pmp->ihidden = hammer2_inode_create(pmp->iroot, NULL, NULL,
1180                                                     NULL, 0,
1181                                 /* lhc */           HAMMER2_INODE_HIDDENDIR,
1182                                 /* inum */          HAMMER2_INODE_HIDDENDIR,
1183                                 /* type */          HAMMER2_OBJTYPE_DIRECTORY,
1184                                 /* target_type */   0,
1185                                 /* flags */         0,
1186                                                     &error);
1187                 if (pmp->ihidden) {
1188                         hammer2_inode_ref(pmp->ihidden);
1189                         hammer2_inode_unlock(pmp->ihidden);
1190                 }
1191                 if (error)
1192                         kprintf("PFS CREATE ERROR %d\n", error);
1193         }
1194
1195         /*
1196          * Scan the hidden directory on-mount and destroy its contents
1197          */
1198         if (error == 0) {
1199                 hammer2_xop_unlinkall_t *xop;
1200
1201                 hammer2_inode_lock(pmp->ihidden, 0);
1202                 xop = hammer2_xop_alloc(pmp->ihidden, HAMMER2_XOP_MODIFYING);
1203                 xop->key_beg = HAMMER2_KEY_MIN;
1204                 xop->key_end = HAMMER2_KEY_MAX;
1205                 hammer2_xop_start(&xop->head, hammer2_inode_xop_unlinkall);
1206
1207                 while ((error = hammer2_xop_collect(&xop->head, 0)) == 0) {
1208                         ;
1209                 }
1210                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1211                 hammer2_inode_unlock(pmp->ihidden);
1212         }
1213
1214         hammer2_inode_unlock(pmp->iroot);
1215         hammer2_trans_done(pmp);
1216 }
1217
1218 /*
1219  * Find the directory common to both fdip and tdip.
1220  *
1221  * Returns a held but not locked inode.  Caller typically locks the inode,
1222  * and when through unlocks AND drops it.
1223  */
1224 hammer2_inode_t *
1225 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1226 {
1227         hammer2_inode_t *scan1;
1228         hammer2_inode_t *scan2;
1229
1230         /*
1231          * We used to have a depth field but it complicated matters too
1232          * much for directory renames.  So now its ugly.  Check for
1233          * simple cases before giving up and doing it the expensive way.
1234          *
1235          * XXX need a bottom-up topology stability lock
1236          */
1237         if (fdip == tdip || fdip == tdip->pip) {
1238                 hammer2_inode_ref(fdip);
1239                 return(fdip);
1240         }
1241         if (fdip->pip == tdip) {
1242                 hammer2_inode_ref(tdip);
1243                 return(tdip);
1244         }
1245
1246         /*
1247          * XXX not MPSAFE
1248          */
1249         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1250                 scan2 = tdip;
1251                 while (scan2->pmp == tdip->pmp) {
1252                         if (scan1 == scan2) {
1253                                 hammer2_inode_ref(scan1);
1254                                 return(scan1);
1255                         }
1256                         scan2 = scan2->pip;
1257                         if (scan2 == NULL)
1258                                 break;
1259                 }
1260         }
1261         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1262               fdip, tdip);
1263         /* NOT REACHED */
1264         return(NULL);
1265 }
1266
1267 /*
1268  * Mark an inode as being modified, meaning that the caller will modify
1269  * ip->meta.
1270  *
1271  * NOTE: No mtid (modify_tid) is passed into this routine.  The caller is
1272  *       only modifying the in-memory inode.  A modify_tid is synchronized
1273  *       later when the inode gets flushed.
1274  */
1275 void
1276 hammer2_inode_modify(hammer2_inode_t *ip)
1277 {
1278         atomic_set_int(&ip->flags, HAMMER2_INODE_MODIFIED);
1279         if (ip->vp)
1280                 vsetisdirty(ip->vp);
1281 }
1282
1283 /*
1284  * Synchronize the inode's frontend state with the chain state prior
1285  * to any explicit flush of the inode or any strategy write call.
1286  *
1287  * Called with a locked inode inside a transaction.
1288  */
1289 void
1290 hammer2_inode_fsync(hammer2_inode_t *ip)
1291 {
1292         if (ip->flags & (HAMMER2_INODE_RESIZED | HAMMER2_INODE_MODIFIED)) {
1293                 hammer2_xop_fsync_t *xop;
1294                 int error;
1295
1296                 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
1297                 xop->clear_directdata = 0;
1298                 if (ip->flags & HAMMER2_INODE_RESIZED) {
1299                         if ((ip->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1300                             ip->meta.size > HAMMER2_EMBEDDED_BYTES) {
1301                                 ip->meta.op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1302                                 xop->clear_directdata = 1;
1303                         }
1304                         xop->osize = ip->osize;
1305                 } else {
1306                         xop->osize = ip->meta.size;     /* safety */
1307                 }
1308                 xop->ipflags = ip->flags;
1309                 xop->meta = ip->meta;
1310
1311                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED |
1312                                              HAMMER2_INODE_MODIFIED);
1313                 hammer2_xop_start(&xop->head, hammer2_inode_xop_fsync);
1314                 error = hammer2_xop_collect(&xop->head, 0);
1315                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1316                 if (error == ENOENT)
1317                         error = 0;
1318                 if (error) {
1319                         kprintf("hammer2: unable to fsync inode %p\n", ip);
1320                         /*
1321                         atomic_set_int(&ip->flags,
1322                                        xop->ipflags & (HAMMER2_INODE_RESIZED |
1323                                                        HAMMER2_INODE_MODIFIED));
1324                         */
1325                         /* XXX return error somehow? */
1326                 }
1327         }
1328 }
1329
1330 /*
1331  * This handles unlinked open files after the vnode is finally dereferenced.
1332  * To avoid deadlocks it cannot be called from the normal vnode recycling
1333  * path, so we call it (1) after a unlink, rmdir, or rename, (2) on every
1334  * flush, and (3) on umount.
1335  *
1336  * Caller must be in a transaction.
1337  */
1338 void
1339 hammer2_inode_run_unlinkq(hammer2_pfs_t *pmp)
1340 {
1341         hammer2_xop_destroy_t *xop;
1342         hammer2_inode_unlink_t *ipul;
1343         hammer2_inode_t *ip;
1344         int error;
1345
1346         if (TAILQ_EMPTY(&pmp->unlinkq))
1347                 return;
1348
1349         LOCKSTART;
1350         hammer2_spin_ex(&pmp->list_spin);
1351         while ((ipul = TAILQ_FIRST(&pmp->unlinkq)) != NULL) {
1352                 TAILQ_REMOVE(&pmp->unlinkq, ipul, entry);
1353                 hammer2_spin_unex(&pmp->list_spin);
1354                 ip = ipul->ip;
1355                 kfree(ipul, pmp->minode);
1356
1357                 hammer2_inode_lock(ip, 0);
1358                 xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
1359                 hammer2_xop_start(&xop->head, hammer2_inode_xop_destroy);
1360                 error = hammer2_xop_collect(&xop->head, 0);
1361                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1362
1363                 hammer2_inode_unlock(ip);
1364                 hammer2_inode_drop(ip);                 /* ipul ref */
1365
1366                 hammer2_spin_ex(&pmp->list_spin);
1367         }
1368         hammer2_spin_unex(&pmp->list_spin);
1369         LOCKSTOP;
1370 }
1371
1372 /*
1373  * Inode create helper (threaded, backend)
1374  *
1375  * Used by ncreate, nmknod, nsymlink, nmkdir.
1376  * Used by nlink and rename to create HARDLINK pointers.
1377  *
1378  * Frontend holds the parent directory ip locked exclusively.  We
1379  * create the inode and feed the exclusively locked chain to the
1380  * frontend.
1381  */
1382 void
1383 hammer2_inode_xop_create(hammer2_xop_t *arg, int clindex)
1384 {
1385         hammer2_xop_create_t *xop = &arg->xop_create;
1386         hammer2_chain_t *parent;
1387         hammer2_chain_t *chain;
1388         hammer2_key_t key_next;
1389         int cache_index = -1;
1390         int error;
1391
1392         kprintf("inode_create lhc %016jx clindex %d\n",
1393                 xop->lhc, clindex);
1394
1395         chain = NULL;
1396         parent = hammer2_inode_chain(xop->head.ip, clindex,
1397                                      HAMMER2_RESOLVE_ALWAYS);
1398         if (parent == NULL) {
1399                 error = EIO;
1400                 goto fail;
1401         }
1402         chain = hammer2_chain_lookup(&parent, &key_next,
1403                                      xop->lhc, xop->lhc,
1404                                      &cache_index, 0);
1405         if (chain) {
1406                 hammer2_chain_unlock(chain);
1407                 error = EEXIST;
1408                 goto fail;
1409         }
1410
1411         error = hammer2_chain_create(&parent, &chain,
1412                                      xop->head.ip->pmp,
1413                                      xop->lhc, 0,
1414                                      HAMMER2_BREF_TYPE_INODE,
1415                                      HAMMER2_INODE_BYTES,
1416                                      xop->head.mtid, xop->flags);
1417         if (error == 0) {
1418                 hammer2_chain_modify(chain, xop->head.mtid, 0);
1419                 chain->data->ipdata.meta = xop->meta;
1420                 if (xop->head.name) {
1421                         bcopy(xop->head.name,
1422                               chain->data->ipdata.filename,
1423                               xop->head.name_len);
1424                         chain->data->ipdata.meta.name_len = xop->head.name_len;
1425                 }
1426                 chain->data->ipdata.meta.name_key = xop->lhc;
1427         }
1428         hammer2_chain_unlock(chain);
1429         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
1430                                   HAMMER2_RESOLVE_SHARED);
1431 fail:
1432         if (parent) {
1433                 hammer2_chain_unlock(parent);
1434                 hammer2_chain_drop(parent);
1435         }
1436         hammer2_xop_feed(&xop->head, chain, clindex, error);
1437         if (chain)
1438                 hammer2_chain_drop(chain);
1439 }
1440
1441 /*
1442  * Inode delete helper (backend, threaded)
1443  *
1444  * Generally used by hammer2_run_unlinkq()
1445  */
1446 void
1447 hammer2_inode_xop_destroy(hammer2_xop_t *arg, int clindex)
1448 {
1449         hammer2_xop_destroy_t *xop = &arg->xop_destroy;
1450         hammer2_pfs_t *pmp;
1451         hammer2_chain_t *parent;
1452         hammer2_chain_t *chain;
1453         hammer2_inode_t *ip;
1454         int error;
1455
1456         /*
1457          * We need the precise parent chain to issue the deletion.
1458          */
1459         ip = xop->head.ip;
1460         pmp = ip->pmp;
1461         chain = NULL;
1462
1463         parent = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
1464         if (parent)
1465                 hammer2_chain_getparent(&parent, HAMMER2_RESOLVE_ALWAYS);
1466         if (parent == NULL) {
1467                 error = EIO;
1468                 goto done;
1469         }
1470         chain = hammer2_inode_chain(ip, clindex, HAMMER2_RESOLVE_ALWAYS);
1471         if (chain == NULL) {
1472                 error = EIO;
1473                 goto done;
1474         }
1475         hammer2_chain_delete(parent, chain, xop->head.mtid, 0);
1476         error = 0;
1477 done:
1478         hammer2_xop_feed(&xop->head, NULL, clindex, error);
1479         if (parent) {
1480                 hammer2_chain_unlock(parent);
1481                 hammer2_chain_drop(parent);
1482         }
1483         if (chain) {
1484                 hammer2_chain_unlock(chain);
1485                 hammer2_chain_drop(chain);
1486         }
1487 }
1488
1489 void
1490 hammer2_inode_xop_unlinkall(hammer2_xop_t *arg, int clindex)
1491 {
1492         hammer2_xop_unlinkall_t *xop = &arg->xop_unlinkall;
1493         hammer2_chain_t *parent;
1494         hammer2_chain_t *chain;
1495         hammer2_key_t key_next;
1496         int cache_index = -1;
1497
1498         /*
1499          * We need the precise parent chain to issue the deletion.
1500          */
1501         parent = hammer2_inode_chain(xop->head.ip, clindex,
1502                                      HAMMER2_RESOLVE_ALWAYS);
1503         chain = NULL;
1504         if (parent == NULL) {
1505                 /* XXX error */
1506                 goto done;
1507         }
1508         chain = hammer2_chain_lookup(&parent, &key_next,
1509                                      xop->key_beg, xop->key_end,
1510                                      &cache_index,
1511                                      HAMMER2_LOOKUP_ALWAYS);
1512         while (chain) {
1513                 hammer2_chain_delete(parent, chain,
1514                                      xop->head.mtid, HAMMER2_DELETE_PERMANENT);
1515                 hammer2_chain_unlock(chain);
1516                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
1517                                           HAMMER2_RESOLVE_SHARED);
1518                 hammer2_xop_feed(&xop->head, chain, clindex, chain->error);
1519                 chain = hammer2_chain_next(&parent, chain, &key_next,
1520                                            key_next, xop->key_end,
1521                                            &cache_index,
1522                                            HAMMER2_LOOKUP_ALWAYS |
1523                                            HAMMER2_LOOKUP_NOUNLOCK);
1524         }
1525 done:
1526         hammer2_xop_feed(&xop->head, NULL, clindex, ENOENT);
1527         if (parent) {
1528                 hammer2_chain_unlock(parent);
1529                 hammer2_chain_drop(parent);
1530         }
1531         if (chain) {
1532                 hammer2_chain_unlock(chain);
1533                 hammer2_chain_drop(chain);
1534         }
1535 }
1536
1537 void
1538 hammer2_inode_xop_connect(hammer2_xop_t *arg, int clindex)
1539 {
1540         hammer2_xop_connect_t *xop = &arg->xop_connect;
1541         hammer2_inode_data_t *wipdata;
1542         hammer2_chain_t *parent;
1543         hammer2_chain_t *chain;
1544         hammer2_pfs_t *pmp;
1545         hammer2_key_t key_dummy;
1546         int cache_index = -1;
1547         int error;
1548
1549         /*
1550          * Get directory, then issue a lookup to prime the parent chain
1551          * for the create.  The lookup is expected to fail.
1552          */
1553         pmp = xop->head.ip->pmp;
1554         parent = hammer2_inode_chain(xop->head.ip, clindex,
1555                                      HAMMER2_RESOLVE_ALWAYS);
1556         if (parent == NULL) {
1557                 chain = NULL;
1558                 error = EIO;
1559                 goto fail;
1560         }
1561         chain = hammer2_chain_lookup(&parent, &key_dummy,
1562                                      xop->lhc, xop->lhc,
1563                                      &cache_index, 0);
1564         if (chain) {
1565                 hammer2_chain_unlock(chain);
1566                 hammer2_chain_drop(chain);
1567                 chain = NULL;
1568                 error = EEXIST;
1569                 goto fail;
1570         }
1571
1572         /*
1573          * Adjust the filename in the inode, set the name key.
1574          *
1575          * NOTE: Frontend must also adjust ip2->meta on success, we can't
1576          *       do it here.
1577          */
1578         chain = hammer2_inode_chain(xop->head.ip2, clindex,
1579                                     HAMMER2_RESOLVE_ALWAYS);
1580         hammer2_chain_modify(chain, xop->head.mtid, 0);
1581         wipdata = &chain->data->ipdata;
1582
1583         hammer2_inode_modify(xop->head.ip2);
1584         if (xop->head.name) {
1585                 bzero(wipdata->filename, sizeof(wipdata->filename));
1586                 bcopy(xop->head.name, wipdata->filename, xop->head.name_len);
1587                 wipdata->meta.name_len = xop->head.name_len;
1588         }
1589         wipdata->meta.name_key = xop->lhc;
1590
1591         /*
1592          * Reconnect the chain to the new parent directory
1593          */
1594         error = hammer2_chain_create(&parent, &chain, pmp,
1595                                      xop->lhc, 0,
1596                                      HAMMER2_BREF_TYPE_INODE,
1597                                      HAMMER2_INODE_BYTES,
1598                                      xop->head.mtid, 0);
1599
1600         /*
1601          * Feed result back.
1602          */
1603 fail:
1604         hammer2_xop_feed(&xop->head, NULL, clindex, error);
1605         if (parent) {
1606                 hammer2_chain_unlock(parent);
1607                 hammer2_chain_drop(parent);
1608         }
1609         if (chain) {
1610                 hammer2_chain_unlock(chain);
1611                 hammer2_chain_drop(chain);
1612         }
1613 }
1614
1615 void
1616 hammer2_inode_xop_fsync(hammer2_xop_t *arg, int clindex)
1617 {
1618         hammer2_xop_fsync_t *xop = &arg->xop_fsync;
1619         hammer2_chain_t *parent;
1620         hammer2_chain_t *chain;
1621         int error;
1622
1623         parent = hammer2_inode_chain(xop->head.ip, clindex,
1624                                      HAMMER2_RESOLVE_ALWAYS);
1625         chain = NULL;
1626         if (parent == NULL) {
1627                 error = EIO;
1628                 goto done;
1629         }
1630         if (parent->error) {
1631                 error = parent->error;
1632                 goto done;
1633         }
1634
1635         error = 0;
1636
1637         if ((xop->ipflags & HAMMER2_INODE_RESIZED) == 0) {
1638                 /* osize must be ignored */
1639         } else if (xop->meta.size < xop->osize) {
1640                 /*
1641                  * We must delete any chains beyond the EOF.  The chain
1642                  * straddling the EOF will be pending in the bioq.
1643                  */
1644                 hammer2_key_t lbase;
1645                 hammer2_key_t key_next;
1646                 int cache_index = -1;
1647
1648                 lbase = (xop->meta.size + HAMMER2_PBUFMASK64) &
1649                         ~HAMMER2_PBUFMASK64;
1650                 chain = hammer2_chain_lookup(&parent, &key_next,
1651                                              lbase, HAMMER2_KEY_MAX,
1652                                              &cache_index,
1653                                              HAMMER2_LOOKUP_NODATA |
1654                                              HAMMER2_LOOKUP_NODIRECT);
1655                 while (chain) {
1656                         /*
1657                          * Degenerate embedded case, nothing to loop on
1658                          */
1659                         switch (chain->bref.type) {
1660                         case HAMMER2_BREF_TYPE_INODE:
1661                                 KKASSERT(0);
1662                                 break;
1663                         case HAMMER2_BREF_TYPE_DATA:
1664                                 hammer2_chain_delete(parent, chain,
1665                                                      xop->head.mtid,
1666                                                      HAMMER2_DELETE_PERMANENT);
1667                                 break;
1668                         }
1669                         chain = hammer2_chain_next(&parent, chain, &key_next,
1670                                                    key_next, HAMMER2_KEY_MAX,
1671                                                    &cache_index,
1672                                                    HAMMER2_LOOKUP_NODATA |
1673                                                    HAMMER2_LOOKUP_NODIRECT);
1674                 }
1675         }
1676
1677         /*
1678          * Sync the inode meta-data, potentially clear the blockset area
1679          * of direct data so it can be used for blockrefs.
1680          */
1681         hammer2_chain_modify(parent, xop->head.mtid, 0);
1682         parent->data->ipdata.meta = xop->meta;
1683         if (xop->clear_directdata) {
1684                 bzero(&parent->data->ipdata.u.blockset,
1685                       sizeof(parent->data->ipdata.u.blockset));
1686         }
1687 done:
1688         if (chain) {
1689                 hammer2_chain_unlock(chain);
1690                 hammer2_chain_drop(chain);
1691         }
1692         if (parent) {
1693                 hammer2_chain_unlock(parent);
1694                 hammer2_chain_drop(parent);
1695         }
1696         hammer2_xop_feed(&xop->head, NULL, clindex, error);
1697 }
1698