sys/vfs/hammer2/hammer2_inode.c

   1 /*
   2  * Copyright (c) 2011-2013 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #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 RB_GENERATE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
  45              hammer2_tid_t, inum);
  46
  47 int
  48 hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2)
  49 {
  50         if (ip1->inum < ip2->inum)
  51                 return(-1);
  52         if (ip1->inum > ip2->inum)
  53                 return(1);
  54         return(0);
  55 }
  56
  57 hammer2_inode_t *
  58 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
  59 {
  60         hammer2_inode_t *ip;
  61
  62         if (pmp) {
  63                 spin_lock(&pmp->inum_spin);
  64                 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
  65                 if (ip)
  66                         hammer2_inode_ref(ip);
  67                 spin_unlock(&pmp->inum_spin);
  68         } else {
  69                 ip = NULL;
  70         }
  71         return(ip);
  72 }
  73
  74 /*
  75  * Adding a ref to an inode is only legal if the inode already has at least
  76  * one ref.
  77  */
  78 void
  79 hammer2_inode_ref(hammer2_inode_t *ip)
  80 {
  81         atomic_add_int(&ip->refs, 1);
  82 }
  83
  84 /*
  85  * Drop an inode reference, freeing the inode when the last reference goes
  86  * away.
  87  */
  88 void
  89 hammer2_inode_drop(hammer2_inode_t *ip)
  90 {
  91         hammer2_mount_t *hmp;
  92         hammer2_inode_t *pip;
  93         hammer2_chain_t *chain;
  94         u_int refs;
  95
  96         while (ip) {
  97                 refs = ip->refs;
  98                 cpu_ccfence();
  99                 if (refs == 1) {
 100                         /*
 101                          * Transition to zero, must interlock with
 102                          * the inode inumber lookup tree (if applicable).
 103                          *
 104                          * NOTE: The super-root inode has no pmp.
 105                          */
 106                         if (ip->pmp)
 107                                 spin_lock(&ip->pmp->inum_spin);
 108
 109                         if (atomic_cmpset_int(&ip->refs, 1, 0)) {
 110                                 KKASSERT(ip->topo_cst.count == 0);
 111                                 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
 112                                         atomic_clear_int(&ip->flags,
 113                                                      HAMMER2_INODE_ONRBTREE);
 114                                         RB_REMOVE(hammer2_inode_tree,
 115                                                   &ip->pmp->inum_tree,
 116                                                   ip);
 117                                 }
 118                                 if (ip->pmp)
 119                                         spin_unlock(&ip->pmp->inum_spin);
 120
 121                                 hmp = ip->hmp;
 122                                 ip->hmp = NULL;
 123                                 pip = ip->pip;
 124                                 ip->pip = NULL;
 125                                 chain = ip->chain;
 126                                 ip->chain = NULL;
 127                                 if (chain)
 128                                         hammer2_chain_drop(chain);
 129
 130                                 /*
 131                                  * We have to drop pip (if non-NULL) to
 132                                  * dispose of our implied reference from
 133                                  * ip->pip.  We can simply loop on it.
 134                                  */
 135                                 kfree(ip, hmp->minode);
 136                                 ip = pip;
 137                                 /* continue with pip (can be NULL) */
 138                         } else {
 139                                 if (ip->pmp)
 140                                         spin_unlock(&ip->pmp->inum_spin);
 141                         }
 142                 } else {
 143                         /*
 144                          * Non zero transition
 145                          */
 146                         if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
 147                                 break;
 148                 }
 149         }
 150 }
 151
 152 /*
 153  * Get the vnode associated with the given inode, allocating the vnode if
 154  * necessary.  The vnode will be returned exclusively locked.
 155  *
 156  * The caller must lock the inode (shared or exclusive).
 157  *
 158  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
 159  * races.
 160  */
 161 struct vnode *
 162 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
 163 {
 164         hammer2_inode_data_t *ipdata;
 165         hammer2_pfsmount_t *pmp;
 166         struct vnode *vp;
 167         ccms_state_t ostate;
 168
 169         pmp = ip->pmp;
 170         KKASSERT(pmp != NULL);
 171         *errorp = 0;
 172         ipdata = &ip->chain->data->ipdata;
 173
 174         for (;;) {
 175                 /*
 176                  * Attempt to reuse an existing vnode assignment.  It is
 177                  * possible to race a reclaim so the vget() may fail.  The
 178                  * inode must be unlocked during the vget() to avoid a
 179                  * deadlock against a reclaim.
 180                  */
 181                 vp = ip->vp;
 182                 if (vp) {
 183                         /*
 184                          * Inode must be unlocked during the vget() to avoid
 185                          * possible deadlocks, but leave the ip ref intact.
 186                          *
 187                          * vnode is held to prevent destruction during the
 188                          * vget().  The vget() can still fail if we lost
 189                          * a reclaim race on the vnode.
 190                          */
 191                         vhold_interlocked(vp);
 192                         ostate = hammer2_inode_lock_temp_release(ip);
 193                         if (vget(vp, LK_EXCLUSIVE)) {
 194                                 vdrop(vp);
 195                                 hammer2_inode_lock_temp_restore(ip, ostate);
 196                                 continue;
 197                         }
 198                         hammer2_inode_lock_temp_restore(ip, ostate);
 199                         vdrop(vp);
 200                         /* vp still locked and ref from vget */
 201                         if (ip->vp != vp) {
 202                                 kprintf("hammer2: igetv race %p/%p\n",
 203                                         ip->vp, vp);
 204                                 vput(vp);
 205                                 continue;
 206                         }
 207                         *errorp = 0;
 208                         break;
 209                 }
 210
 211                 /*
 212                  * No vnode exists, allocate a new vnode.  Beware of
 213                  * allocation races.  This function will return an
 214                  * exclusively locked and referenced vnode.
 215                  */
 216                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
 217                 if (*errorp) {
 218                         kprintf("hammer2: igetv getnewvnode failed %d\n",
 219                                 *errorp);
 220                         vp = NULL;
 221                         break;
 222                 }
 223
 224                 /*
 225                  * Lock the inode and check for an allocation race.
 226                  */
 227                 ostate = hammer2_inode_lock_upgrade(ip);
 228                 if (ip->vp != NULL) {
 229                         vp->v_type = VBAD;
 230                         vx_put(vp);
 231                         hammer2_inode_lock_downgrade(ip, ostate);
 232                         continue;
 233                 }
 234
 235                 switch (ipdata->type) {
 236                 case HAMMER2_OBJTYPE_DIRECTORY:
 237                         vp->v_type = VDIR;
 238                         break;
 239                 case HAMMER2_OBJTYPE_REGFILE:
 240                         vp->v_type = VREG;
 241                         vinitvmio(vp, ipdata->size,
 242                                   HAMMER2_LBUFSIZE,
 243                                   (int)ipdata->size & HAMMER2_LBUFMASK);
 244                         break;
 245                 case HAMMER2_OBJTYPE_SOFTLINK:
 246                         /*
 247                          * XXX for now we are using the generic file_read
 248                          * and file_write code so we need a buffer cache
 249                          * association.
 250                          */
 251                         vp->v_type = VLNK;
 252                         vinitvmio(vp, ipdata->size,
 253                                   HAMMER2_LBUFSIZE,
 254                                   (int)ipdata->size & HAMMER2_LBUFMASK);
 255                         break;
 256                 /* XXX FIFO */
 257                 default:
 258                         panic("hammer2: unhandled objtype %d", ipdata->type);
 259                         break;
 260                 }
 261
 262                 if (ip == pmp->iroot)
 263                         vsetflags(vp, VROOT);
 264
 265                 vp->v_data = ip;
 266                 ip->vp = vp;
 267                 hammer2_inode_ref(ip);          /* vp association */
 268                 hammer2_inode_lock_downgrade(ip, ostate);
 269                 break;
 270         }
 271
 272         /*
 273          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
 274          */
 275         if (hammer2_debug & 0x0002) {
 276                 kprintf("igetv vp %p refs %d aux %d\n",
 277                         vp, vp->v_sysref.refcnt, vp->v_auxrefs);
 278         }
 279         return (vp);
 280 }
 281
 282 /*
 283  * The passed-in chain must be locked and the returned inode will also be
 284  * locked.  This routine typically locates or allocates the inode, assigns
 285  * ip->chain (adding a ref to chain if necessary), and returns the inode.
 286  *
 287  * The hammer2_inode structure regulates the interface between the high level
 288  * kernel VNOPS API and the filesystem backend (the chains).
 289  *
 290  * WARNING!  This routine sucks up the chain's lock (makes it part of the
 291  *           inode lock from the point of view of the inode lock API),
 292  *           so callers need to be careful.
 293  *
 294  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
 295  *           is allowed to pass pmp == NULL and dip == NULL for sroot.
 296  */
 297 hammer2_inode_t *
 298 hammer2_inode_get(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
 299                   hammer2_inode_t *dip, hammer2_chain_t *chain)
 300 {
 301         hammer2_inode_t *nip;
 302         hammer2_chain_t *ochain;
 303
 304         KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
 305
 306         /*
 307          * Interlocked lookup/ref of the inode.  This code is only needed
 308          * when looking up inodes with nlinks != 0 (TODO: optimize out
 309          * otherwise and test for duplicates).
 310          */
 311 again:
 312         for (;;) {
 313                 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
 314                 if (nip == NULL)
 315                         break;
 316                 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
 317                 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) {
 318                         ccms_thread_unlock(&nip->topo_cst);
 319                         hammer2_inode_drop(nip);
 320                         continue;
 321                 }
 322                 if (nip->chain != chain) {
 323                         hammer2_chain_ref(chain);       /* new nip->chain */
 324                         ochain = nip->chain;
 325                         nip->chain = chain;             /* fully locked   */
 326                         hammer2_chain_drop(ochain);     /* old nip->chain */
 327                 }
 328
 329                 /*
 330                  * Consolidated nip/nip->chain is locked (chain locked
 331                  * by caller).
 332                  */
 333                 return nip;
 334         }
 335
 336         /*
 337          * We couldn't find the inode number, create a new inode.
 338          */
 339         nip = kmalloc(sizeof(*nip), hmp->minode, M_WAITOK | M_ZERO);
 340         nip->inum = chain->data->ipdata.inum;
 341         nip->chain = chain;
 342         hammer2_chain_ref(chain);               /* nip->chain */
 343         nip->pip = dip;                         /* can be NULL */
 344         if (dip)
 345                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
 346
 347         nip->pmp = pmp;
 348         nip->hmp = hmp;
 349
 350         /*
 351          * ref and lock on nip gives it state compatible to after a
 352          * hammer2_inode_lock_ex() call.
 353          */
 354         nip->refs = 1;
 355         ccms_cst_init(&nip->topo_cst, &nip->chain);
 356         ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
 357         /* combination of thread lock and chain lock == inode lock */
 358
 359         /*
 360          * Attempt to add the inode.  If it fails we raced another inode
 361          * get.  Undo all the work and try again.
 362          */
 363         if (pmp) {
 364                 spin_lock(&pmp->inum_spin);
 365                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
 366                         spin_unlock(&pmp->inum_spin);
 367                         ccms_thread_unlock(&nip->topo_cst);
 368                         hammer2_inode_drop(nip);
 369                         goto again;
 370                 }
 371                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
 372                 spin_unlock(&pmp->inum_spin);
 373         }
 374
 375         return (nip);
 376 }
 377
 378 /*
 379  * Put away an inode, unlocking it and disconnecting it from its chain.
 380  *
 381  * The inode must be exclusively locked on call and non-recursed, with
 382  * at least 2 refs (one belonging to the exclusive lock, and one additional
 383  * ref belonging to the caller).
 384  *
 385  * Upon return the inode typically has one ref remaining which the caller
 386  * drops.
 387  */
 388 void
 389 hammer2_inode_put(hammer2_inode_t *ip, hammer2_chain_t *chain)
 390 {
 391         hammer2_inode_t *pip;
 392
 393         /*
 394          * Disconnect and unlock chain
 395          */
 396         KKASSERT(ip->refs >= 2);
 397         KKASSERT(ip->topo_cst.count == -1);     /* one excl lock allowed */
 398         if ((chain = ip->chain) != NULL) {
 399                 ip->chain = NULL;
 400                 hammer2_chain_drop(chain);      /* from *_get() */
 401         }
 402
 403         /*
 404          * Disconnect pip
 405          */
 406         pip = ip->pip;
 407         ip->pip = NULL;
 408         hammer2_inode_unlock_ex(ip, chain);
 409
 410         /*
 411          * Cleanup delayed actions
 412          */
 413         if (pip)
 414                 hammer2_inode_drop(pip);
 415 }
 416
 417 /*
 418  * Create a new inode in the specified directory using the vattr to
 419  * figure out the type of inode.
 420  *
 421  * If no error occurs the new inode with its chain locked is returned in
 422  * *nipp, otherwise an error is returned and *nipp is set to NULL.
 423  *
 424  * If vap and/or cred are NULL the related fields are not set and the
 425  * inode type defaults to a directory.  This is used when creating PFSs
 426  * under the super-root, so the inode number is set to 1 in this case.
 427  *
 428  * dip is not locked on entry.
 429  */
 430 hammer2_inode_t *
 431 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
 432                      struct vattr *vap, struct ucred *cred,
 433                      const uint8_t *name, size_t name_len,
 434                      hammer2_chain_t **chainp, int *errorp)
 435 {
 436         hammer2_inode_data_t *dipdata;
 437         hammer2_inode_data_t *nipdata;
 438         hammer2_mount_t *hmp;
 439         hammer2_chain_t *chain;
 440         hammer2_chain_t *parent;
 441         hammer2_inode_t *nip;
 442         hammer2_key_t lhc;
 443         int error;
 444         uid_t xuid;
 445         uuid_t dip_uid;
 446         uuid_t dip_gid;
 447         uint32_t dip_mode;
 448
 449         hmp = dip->hmp;
 450         lhc = hammer2_dirhash(name, name_len);
 451         *errorp = 0;
 452
 453         /*
 454          * Locate the inode or indirect block to create the new
 455          * entry in.  At the same time check for key collisions
 456          * and iterate until we don't get one.
 457          *
 458          * NOTE: hidden inodes do not have iterators.
 459          */
 460 retry:
 461         parent = hammer2_inode_lock_ex(dip);
 462         dipdata = &dip->chain->data->ipdata;
 463         dip_uid = dipdata->uid;
 464         dip_gid = dipdata->gid;
 465         dip_mode = dipdata->mode;
 466
 467         error = 0;
 468         while (error == 0) {
 469                 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
 470                 if (chain == NULL)
 471                         break;
 472                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
 473                         error = ENOSPC;
 474                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 475                         error = ENOSPC;
 476                 hammer2_chain_unlock(chain);
 477                 chain = NULL;
 478                 ++lhc;
 479         }
 480         if (error == 0) {
 481                 error = hammer2_chain_create(trans, &parent, &chain,
 482                                              lhc, 0,
 483                                              HAMMER2_BREF_TYPE_INODE,
 484                                              HAMMER2_INODE_BYTES);
 485         }
 486
 487         /*
 488          * Cleanup and handle retries.
 489          */
 490         if (error == EAGAIN) {
 491                 hammer2_chain_ref(parent);
 492                 hammer2_inode_unlock_ex(dip, parent);
 493                 hammer2_chain_wait(parent);
 494                 hammer2_chain_drop(parent);
 495                 goto retry;
 496         }
 497         hammer2_inode_unlock_ex(dip, parent);
 498
 499         if (error) {
 500                 KKASSERT(chain == NULL);
 501                 *errorp = error;
 502                 return (NULL);
 503         }
 504
 505         /*
 506          * Set up the new inode.
 507          *
 508          * NOTE: *_get() integrates chain's lock into the inode lock.
 509          *
 510          * NOTE: Only one new inode can currently be created per
 511          *       transaction.  If the need arises we can adjust
 512          *       hammer2_trans_init() to allow more.
 513          */
 514         chain->data->ipdata.inum = trans->sync_tid;
 515         nip = hammer2_inode_get(dip->hmp, dip->pmp, dip, chain);
 516         nipdata = &chain->data->ipdata;
 517
 518         if (vap) {
 519                 KKASSERT(trans->inodes_created == 0);
 520                 nipdata->type = hammer2_get_obj_type(vap->va_type);
 521                 nipdata->inum = trans->sync_tid;
 522                 ++trans->inodes_created;
 523         } else {
 524                 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
 525                 nipdata->inum = 1;
 526         }
 527         nipdata->version = HAMMER2_INODE_VERSION_ONE;
 528         hammer2_update_time(&nipdata->ctime);
 529         nipdata->mtime = nipdata->ctime;
 530         if (vap)
 531                 nipdata->mode = vap->va_mode;
 532         nipdata->nlinks = 1;
 533         if (vap) {
 534                 if (dip) {
 535                         xuid = hammer2_to_unix_xid(&dip_uid);
 536                         xuid = vop_helper_create_uid(dip->pmp->mp,
 537                                                      dip_mode,
 538                                                      xuid,
 539                                                      cred,
 540                                                      &vap->va_mode);
 541                 } else {
 542                         xuid = 0;
 543                 }
 544                 if (vap->va_vaflags & VA_UID_UUID_VALID)
 545                         nipdata->uid = vap->va_uid_uuid;
 546                 else if (vap->va_uid != (uid_t)VNOVAL)
 547                         hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
 548                 else
 549                         hammer2_guid_to_uuid(&nipdata->uid, xuid);
 550
 551                 if (vap->va_vaflags & VA_GID_UUID_VALID)
 552                         nipdata->gid = vap->va_gid_uuid;
 553                 else if (vap->va_gid != (gid_t)VNOVAL)
 554                         hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
 555                 else if (dip)
 556                         nipdata->gid = dip_gid;
 557         }
 558
 559         /*
 560          * Regular files and softlinks allow a small amount of data to be
 561          * directly embedded in the inode.  This flag will be cleared if
 562          * the size is extended past the embedded limit.
 563          */
 564         if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
 565             nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
 566                 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
 567         }
 568
 569         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 570         bcopy(name, nipdata->filename, name_len);
 571         nipdata->name_key = lhc;
 572         nipdata->name_len = name_len;
 573         *chainp = chain;
 574
 575         return (nip);
 576 }
 577
 578 /*
 579  * chain may have been moved around by the create.
 580  */
 581 static
 582 void
 583 hammer2_chain_refactor(hammer2_chain_t **chainp)
 584 {
 585         hammer2_chain_t *chain = *chainp;
 586         hammer2_chain_t *tmp;
 587
 588         while (chain->duplink && (chain->flags & HAMMER2_CHAIN_DELETED)) {
 589                 tmp = chain->duplink;
 590                 while (tmp->duplink && (tmp->flags & HAMMER2_CHAIN_DELETED))
 591                         tmp = tmp->duplink;
 592                 hammer2_chain_ref(chain);
 593                 hammer2_chain_unlock(chain);
 594                 hammer2_chain_lock(tmp, HAMMER2_RESOLVE_ALWAYS);
 595                 hammer2_chain_drop(chain);
 596                 chain = tmp;
 597                 *chainp = chain;
 598         }
 599 }
 600
 601
 602 /*
 603  * ochain represents the target file inode.  We need to move it to the
 604  * specified common parent directory (dip) and rename it to a special
 605  * invisible "0xINODENUMBER" filename.
 606  *
 607  * We use chain_duplicate and duplicate ochain at the new location,
 608  * renaming it appropriately.  We create a temporary chain and
 609  * then delete it to placemark where the duplicate will go.  Both of
 610  * these use the inode number for (lhc) (the key), generating the
 611  * invisible filename.
 612  */
 613 static
 614 hammer2_chain_t *
 615 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
 616                         hammer2_inode_t *dip, int *errorp)
 617 {
 618         hammer2_inode_data_t *nipdata;
 619         hammer2_mount_t *hmp;
 620         hammer2_chain_t *parent;
 621         hammer2_chain_t *ochain;
 622         hammer2_chain_t *nchain;
 623         hammer2_chain_t *tmp;
 624         hammer2_key_t lhc;
 625         hammer2_blockref_t bref;
 626
 627         ochain = *ochainp;
 628         *errorp = 0;
 629         hmp = dip->hmp;
 630         lhc = ochain->data->ipdata.inum;
 631         KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
 632
 633         /*
 634          * Locate the inode or indirect block to create the new
 635          * entry in.  lhc represents the inode number so there is
 636          * no collision iteration.
 637          *
 638          * There should be no key collisions with invisible inode keys.
 639          */
 640 retry:
 641         parent = hammer2_chain_lookup_init(dip->chain, 0);
 642         nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
 643         if (nchain) {
 644                 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
 645                         nchain, parent, dip, dip->chain);
 646                 hammer2_chain_unlock(nchain);
 647                 nchain = NULL;
 648                 *errorp = ENOSPC;
 649 #if 0
 650                 Debugger("X3");
 651 #endif
 652         }
 653
 654         /*
 655          * Create entry in common parent directory using the seek position
 656          * calculated above.
 657          */
 658         if (*errorp == 0) {
 659                 KKASSERT(nchain == NULL);
 660                 *errorp = hammer2_chain_create(trans, &parent, &nchain,
 661                                                lhc, 0,
 662                                                HAMMER2_BREF_TYPE_INODE,/* n/a */
 663                                                HAMMER2_INODE_BYTES);   /* n/a */
 664                 hammer2_chain_refactor(&ochain);
 665                 *ochainp = ochain;
 666         }
 667
 668         /*
 669          * Cleanup and handle retries.
 670          */
 671         if (*errorp == EAGAIN) {
 672                 hammer2_chain_ref(parent);
 673                 hammer2_chain_lookup_done(parent);
 674                 hammer2_chain_wait(parent);
 675                 hammer2_chain_drop(parent);
 676                 goto retry;
 677         }
 678
 679         /*
 680          * Handle the error case
 681          */
 682         if (*errorp) {
 683                 KKASSERT(nchain == NULL);
 684                 hammer2_chain_lookup_done(parent);
 685                 return (NULL);
 686         }
 687
 688         /*
 689          * Use chain as a placeholder for (lhc), delete it and replace
 690          * it with our duplication.
 691          *
 692          * Gain a second lock on ochain for the duplication function to
 693          * unlock, maintain the caller's original lock across the call.
 694          *
 695          * This is a bit messy.
 696          */
 697         hammer2_chain_delete(trans, parent, nchain);
 698         hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
 699         tmp = ochain;
 700         bref = tmp->bref;
 701         bref.key = lhc;                 /* invisible dir entry key */
 702         bref.keybits = 0;
 703         hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
 704         hammer2_chain_lookup_done(parent);
 705         hammer2_chain_unlock(nchain);   /* no longer needed */
 706
 707         /*
 708          * Now set chain to our duplicate and modify it appropriately.
 709          *
 710          * Directory entries are inodes but this is a hidden hardlink
 711          * target.  The name isn't used but to ease debugging give it
 712          * a name after its inode number.
 713          */
 714         nchain = tmp;
 715         tmp = NULL;     /* safety */
 716
 717         hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
 718         nipdata = &nchain->data->ipdata;
 719         ksnprintf(nipdata->filename, sizeof(nipdata->filename),
 720                   "0x%016jx", (intmax_t)nipdata->inum);
 721         nipdata->name_len = strlen(nipdata->filename);
 722         nipdata->name_key = lhc;
 723
 724         return (nchain);
 725 }
 726
 727 /*
 728  * Connect the target inode represented by (*chainp) to the media topology
 729  * at (dip, name, len).
 730  *
 731  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
 732  * entry instead of connecting (*chainp).
 733  *
 734  * If hlink is FALSE this function uses chain_duplicate() to make a copy
 735  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
 736  * by the caller in this case (e.g. rename).
 737  */
 738 int
 739 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 740                       hammer2_inode_t *dip, hammer2_chain_t **chainp,
 741                       const uint8_t *name, size_t name_len)
 742 {
 743         hammer2_inode_data_t *ipdata;
 744         hammer2_mount_t *hmp;
 745         hammer2_chain_t *nchain;
 746         hammer2_chain_t *parent;
 747         hammer2_chain_t *ochain;
 748         hammer2_key_t lhc;
 749         int error;
 750
 751         hmp = dip->hmp;
 752
 753         ochain = *chainp;
 754
 755         /*
 756          * Since ochain is either disconnected from the topology or represents
 757          * a hardlink terminus which is always a parent of or equal to dip,
 758          * we should be able to safely lock dip->chain for our setup.
 759          */
 760         parent = hammer2_chain_lookup_init(dip->chain, 0);
 761
 762         lhc = hammer2_dirhash(name, name_len);
 763
 764         /*
 765          * Locate the inode or indirect block to create the new
 766          * entry in.  At the same time check for key collisions
 767          * and iterate until we don't get one.
 768          */
 769         error = 0;
 770         while (error == 0) {
 771                 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
 772                 if (nchain == NULL)
 773                         break;
 774                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 775                         error = ENOSPC;
 776                 hammer2_chain_unlock(nchain);
 777                 nchain = NULL;
 778                 ++lhc;
 779         }
 780
 781         if (error == 0) {
 782                 if (hlink) {
 783                         /*
 784                          * Hardlink pointer needed, create totally fresh
 785                          * directory entry.
 786                          */
 787                         KKASSERT(nchain == NULL);
 788                         error = hammer2_chain_create(trans, &parent, &nchain,
 789                                                      lhc, 0,
 790                                                      HAMMER2_BREF_TYPE_INODE,
 791                                                      HAMMER2_INODE_BYTES);
 792                         hammer2_chain_refactor(&ochain);
 793                 } else {
 794                         /*
 795                          * Reconnect the original chain and rename.  Use
 796                          * chain_duplicate().  The caller will likely delete
 797                          * or has already deleted the original chain in
 798                          * this case.
 799                          */
 800                         nchain = ochain;
 801                         ochain = NULL;
 802                         hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
 803                         error = hammer2_chain_create(trans, &parent, &nchain,
 804                                                      lhc, 0,
 805                                                      HAMMER2_BREF_TYPE_INODE,
 806                                                      HAMMER2_INODE_BYTES);
 807                 }
 808         }
 809
 810         /*
 811          * Unlock stuff.
 812          */
 813         KKASSERT(error != EAGAIN);
 814         hammer2_chain_lookup_done(parent);
 815         parent = NULL;
 816
 817         /*
 818          * nchain should be NULL on error, leave ochain (== *chainp) alone.
 819          */
 820         if (error) {
 821                 KKASSERT(nchain == NULL);
 822                 return (error);
 823         }
 824
 825         /*
 826          * Directory entries are inodes so if the name has changed we have
 827          * to update the inode.
 828          *
 829          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
 830          * chain, the caller will access the hardlink via the actual hardlink
 831          * target file and not the hardlink pointer entry, so we must still
 832          * return ochain.
 833          */
 834         if (hlink && hammer2_hardlink_enable >= 0) {
 835                 /*
 836                  * Create the HARDLINK pointer.  oip represents the hardlink
 837                  * target in this situation.
 838                  *
 839                  * We will return ochain (the hardlink target).
 840                  */
 841                 hammer2_chain_modify(trans, &nchain,
 842                                      HAMMER2_MODIFY_ASSERTNOCOPY);
 843                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 844                 ipdata = &nchain->data->ipdata;
 845                 bcopy(name, ipdata->filename, name_len);
 846                 ipdata->name_key = lhc;
 847                 ipdata->name_len = name_len;
 848                 ipdata->target_type = ochain->data->ipdata.type;
 849                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
 850                 ipdata->inum = ochain->data->ipdata.inum;
 851                 ipdata->nlinks = 1;
 852                 hammer2_chain_unlock(nchain);
 853                 nchain = ochain;
 854                 ochain = NULL;
 855         } else if (hlink && hammer2_hardlink_enable < 0) {
 856                 /*
 857                  * Create a snapshot (hardlink fake mode for debugging).
 858                  * (ochain already flushed above so we can just copy the
 859                  * bref XXX).
 860                  *
 861                  * Since this is a snapshot we return nchain in the fake
 862                  * hardlink case.
 863                  */
 864                 hammer2_chain_modify(trans, &nchain,
 865                                      HAMMER2_MODIFY_ASSERTNOCOPY);
 866                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 867                 ipdata = &nchain->data->ipdata;
 868                 *ipdata = ochain->data->ipdata;
 869                 bcopy(name, ipdata->filename, name_len);
 870                 ipdata->name_key = lhc;
 871                 ipdata->name_len = name_len;
 872                 kprintf("created fake hardlink %*.*s\n",
 873                         (int)name_len, (int)name_len, name);
 874         } else {
 875                 /*
 876                  * nchain is a duplicate of ochain at the new location.
 877                  * We must fixup the name stored in oip.  The bref key
 878                  * has already been set up.
 879                  */
 880                 hammer2_chain_modify(trans, &nchain,
 881                                      HAMMER2_MODIFY_ASSERTNOCOPY);
 882                 ipdata = &nchain->data->ipdata;
 883
 884                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 885                 bcopy(name, ipdata->filename, name_len);
 886                 ipdata->name_key = lhc;
 887                 ipdata->name_len = name_len;
 888                 ipdata->nlinks = 1;
 889         }
 890
 891         /*
 892          * We are replacing ochain with nchain, unlock ochain.  In the
 893          * case where ochain is left unchanged the code above sets
 894          * nchain to ochain and ochain to NULL, resulting in a NOP here.
 895          */
 896         if (ochain)
 897                 hammer2_chain_unlock(ochain);
 898         *chainp = nchain;
 899
 900         return (0);
 901 }
 902
 903 /*
 904  * Caller must hold exactly ONE exclusive lock on the inode.  *nchainp
 905  * must be exclusive locked (its own exclusive lock even if it is the
 906  * same as ip->chain).
 907  *
 908  * This function replaces ip->chain.  The exclusive lock on the passed
 909  * nchain is inherited by the inode and the caller becomes responsible
 910  * for unlocking it when the caller unlocks the inode.
 911  *
 912  * ochain was locked by the caller indirectly via the inode lock.  Since
 913  * ip->chain is being repointed, we become responsible for cleaning up
 914  * that lock.
 915  *
 916  * Return *nchainp = NULL as a safety.
 917  */
 918 void
 919 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 920                       hammer2_chain_t *nchain)
 921 {
 922         hammer2_chain_t *ochain;
 923         hammer2_inode_t *opip;
 924
 925         /*
 926          * ip->chain points to the hardlink target, not the hardlink psuedo
 927          * inode.  Do not repoint nchain to the pseudo-node.
 928          */
 929         if (nchain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK)
 930                 return;
 931
 932         /*
 933          * Repoint ip->chain if necessary.
 934          *
 935          * (Inode must be locked exclusively by parent)
 936          */
 937         ochain = ip->chain;
 938         if (ochain != nchain) {
 939                 hammer2_chain_ref(nchain);              /* for ip->chain */
 940                 ip->chain = nchain;
 941                 if (ochain) {
 942                         hammer2_chain_unlock(ochain);
 943                         hammer2_chain_drop(ochain);     /* for ip->chain */
 944                 }
 945                 /* replace locked chain in ip (additional lock) */
 946                 hammer2_chain_lock(nchain, HAMMER2_RESOLVE_ALWAYS);
 947         }
 948         if (ip->pip != pip) {
 949                 opip = ip->pip;
 950                 if (pip)
 951                         hammer2_inode_ref(pip);
 952                 ip->pip = pip;
 953                 if (opip)
 954                         hammer2_inode_drop(opip);
 955         }
 956 }
 957
 958 /*
 959  * Unlink the file from the specified directory inode.  The directory inode
 960  * does not need to be locked.  The caller should pass a non-NULL (ip)
 961  * representing the object being removed only if the related vnode is
 962  * potentially inactive (not referenced in the caller's active path),
 963  * so we can vref/vrele it to trigger the VOP_INACTIVE path and properly
 964  * recycle it.
 965  *
 966  * isdir determines whether a directory/non-directory check should be made.
 967  * No check is made if isdir is set to -1.
 968  *
 969  * NOTE!  This function does not prevent the underlying file from still
 970  *        being used if it has other refs (such as from an inode, or if it's
 971  *        chain is manually held).  However, the caller is responsible for
 972  *        fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
 973  */
 974 int
 975 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 976                     const uint8_t *name, size_t name_len,
 977                     int isdir, int *hlinkp)
 978 {
 979         hammer2_inode_data_t *ipdata;
 980         hammer2_mount_t *hmp;
 981         hammer2_chain_t *parent;
 982         hammer2_chain_t *ochain;
 983         hammer2_chain_t *chain;
 984         hammer2_chain_t *dparent;
 985         hammer2_chain_t *dchain;
 986         hammer2_key_t lhc;
 987         int error;
 988         int parent_ref;
 989         uint8_t type;
 990
 991         parent_ref = 0;
 992         error = 0;
 993         ochain = NULL;
 994         hmp = dip->hmp;
 995         lhc = hammer2_dirhash(name, name_len);
 996
 997         /*
 998          * Search for the filename in the directory
 999          */
1000         if (hlinkp)
1001                 *hlinkp = 0;
1002         parent = hammer2_inode_lock_ex(dip);
1003         chain = hammer2_chain_lookup(&parent,
1004                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1005                                      0);
1006         while (chain) {
1007                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1008                     name_len == chain->data->ipdata.name_len &&
1009                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1010                         break;
1011                 }
1012                 chain = hammer2_chain_next(&parent, chain,
1013                                            lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1014                                            0);
1015         }
1016         hammer2_inode_unlock_ex(dip, NULL);     /* retain parent */
1017
1018         /*
1019          * Not found or wrong type (isdir < 0 disables the type check).
1020          * If a hardlink pointer, type checks use the hardlink target.
1021          */
1022         if (chain == NULL) {
1023                 error = ENOENT;
1024                 goto done;
1025         }
1026         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1027                 if (hlinkp)
1028                         *hlinkp = 1;
1029                 type = chain->data->ipdata.target_type;
1030         }
1031
1032         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1033                 error = ENOTDIR;
1034                 goto done;
1035         }
1036         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1037                 error = EISDIR;
1038                 goto done;
1039         }
1040
1041         /*
1042          * Hardlink must be resolved.  We can't hold parent locked while we
1043          * do this or we could deadlock.
1044          *
1045          * On success chain will be adjusted to point at the hardlink target
1046          * and ochain will point to the hardlink pointer in the original
1047          * directory.  Otherwise chain remains pointing to the original.
1048          */
1049         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1050                 KKASSERT(parent_ref == 0);
1051                 hammer2_chain_unlock(parent);
1052                 parent = NULL;
1053                 error = hammer2_hardlink_find(dip, &chain, &ochain);
1054         }
1055
1056         /*
1057          * If this is a directory the directory must be empty.  However, if
1058          * isdir < 0 we are doing a rename and the directory does not have
1059          * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1060          * and the directory does not have to be empty.
1061          *
1062          * NOTE: We check the full key range here which covers both visible
1063          *       and invisible entries.  Theoretically there should be no
1064          *       invisible (hardlink target) entries if there are no visible
1065          *       entries.
1066          */
1067         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1068                 dparent = hammer2_chain_lookup_init(chain, 0);
1069                 dchain = hammer2_chain_lookup(&dparent,
1070                                               0, (hammer2_key_t)-1,
1071                                               HAMMER2_LOOKUP_NODATA);
1072                 if (dchain) {
1073                         hammer2_chain_unlock(dchain);
1074                         hammer2_chain_lookup_done(dparent);
1075                         error = ENOTEMPTY;
1076                         goto done;
1077                 }
1078                 hammer2_chain_lookup_done(dparent);
1079                 dparent = NULL;
1080                 /* dchain NULL */
1081         }
1082
1083         /*
1084          * Ok, we can now unlink the chain.  We always decrement nlinks even
1085          * if the entry can be deleted in case someone has the file open and
1086          * does an fstat().
1087          *
1088          * The chain itself will no longer be in the on-media topology but
1089          * can still be flushed to the media (e.g. if an open descriptor
1090          * remains).  When the last vnode/ip ref goes away the chain will
1091          * be marked unmodified, avoiding any further (now unnecesary) I/O.
1092          *
1093          * A non-NULL ochain indicates a hardlink.
1094          */
1095         if (ochain) {
1096                 /*
1097                  * Delete the original hardlink pointer.
1098                  *
1099                  * NOTE: parent from above is NULL when ochain != NULL
1100                  *       so we can reuse it.
1101                  */
1102                 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1103                 parent_ref = 1;
1104                 for (;;) {
1105                         parent = ochain->parent;
1106                         hammer2_chain_ref(parent);
1107                         hammer2_chain_unlock(ochain);
1108                         hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1109                         hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1110                         if (ochain->parent == parent)
1111                                 break;
1112                         hammer2_chain_unlock(parent);
1113                         hammer2_chain_drop(parent);
1114                 }
1115
1116                 hammer2_chain_delete(trans, parent, ochain);
1117                 hammer2_chain_unlock(ochain);
1118                 hammer2_chain_unlock(parent);
1119                 hammer2_chain_drop(parent);
1120                 parent = NULL;
1121
1122                 /*
1123                  * Then decrement nlinks on hardlink target, deleting
1124                  * the target when nlinks drops to 0.
1125                  */
1126                 if (chain->data->ipdata.nlinks == 1) {
1127                         dparent = chain->parent;
1128                         hammer2_chain_ref(chain);
1129                         hammer2_chain_unlock(chain);
1130                         hammer2_chain_lock(dparent, HAMMER2_RESOLVE_ALWAYS);
1131                         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
1132                         hammer2_chain_drop(chain);
1133                         hammer2_chain_modify(trans, &chain, 0);
1134                         --chain->data->ipdata.nlinks;
1135                         hammer2_chain_delete(trans, dparent, chain);
1136                         hammer2_chain_unlock(dparent);
1137                 } else {
1138                         hammer2_chain_modify(trans, &chain, 0);
1139                         --chain->data->ipdata.nlinks;
1140                 }
1141         } else {
1142                 /*
1143                  * Otherwise this was not a hardlink and we can just
1144                  * remove the entry and decrement nlinks.
1145                  *
1146                  * NOTE: *_get() integrates chain's lock into the inode lock.
1147                  */
1148                 hammer2_chain_modify(trans, &chain, 0);
1149                 ipdata = &chain->data->ipdata;
1150                 --ipdata->nlinks;
1151                 hammer2_chain_delete(trans, parent, chain);
1152         }
1153
1154         error = 0;
1155 done:
1156         if (chain)
1157                 hammer2_chain_unlock(chain);
1158         if (parent) {
1159                 hammer2_chain_lookup_done(parent);
1160                 if (parent_ref)
1161                         hammer2_chain_drop(parent);
1162         }
1163         if (ochain)
1164                 hammer2_chain_drop(ochain);
1165
1166         return error;
1167 }
1168
1169 /*
1170  * Calculate the allocation size for the file fragment straddling EOF
1171  */
1172 int
1173 hammer2_inode_calc_alloc(hammer2_key_t filesize)
1174 {
1175         int frag = (int)filesize & HAMMER2_PBUFMASK;
1176         int radix;
1177
1178         if (frag == 0)
1179                 return(0);
1180         for (radix = HAMMER2_MINALLOCRADIX; frag > (1 << radix); ++radix)
1181                 ;
1182         return (radix);
1183 }
1184
1185 /*
1186  * Given an exclusively locked inode we consolidate its chain for hardlink
1187  * creation, adding (nlinks) to the file's link count and potentially
1188  * relocating the inode to a directory common to ip->pip and tdip.
1189  *
1190  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1191  * and returning a new locked chain.
1192  *
1193  * NOTE!  This function will also replace ip->chain.
1194  */
1195 int
1196 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1197                              hammer2_chain_t **chainp,
1198                              hammer2_inode_t *tdip, int nlinks)
1199 {
1200         hammer2_inode_data_t *ipdata;
1201         hammer2_mount_t *hmp;
1202         hammer2_inode_t *fdip;
1203         hammer2_inode_t *cdip;
1204         hammer2_chain_t *chain;
1205         hammer2_chain_t *nchain;
1206         hammer2_chain_t *parent;
1207         int error;
1208
1209         hmp = tdip->hmp;
1210
1211         chain = *chainp;
1212         if (nlinks == 0 &&                      /* no hardlink needed */
1213             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1214                 return (0);
1215         }
1216         if (hammer2_hardlink_enable < 0) {      /* fake hardlinks */
1217                 return (0);
1218         }
1219
1220         if (hammer2_hardlink_enable == 0) {     /* disallow hardlinks */
1221                 hammer2_chain_unlock(chain);
1222                 *chainp = NULL;
1223                 return (ENOTSUP);
1224         }
1225
1226         /*
1227          * cdip will be returned with a ref, but not locked.
1228          */
1229         fdip = ip->pip;
1230         cdip = hammer2_inode_common_parent(fdip, tdip);
1231
1232         /*
1233          * If no change in the hardlink's target directory is required and
1234          * this is already a hardlink target, all we need to do is adjust
1235          * the link count.
1236          *
1237          * XXX The common parent is a big wiggly due to duplication from
1238          *     renames.  Compare the core (RBTREE) pointer instead of the
1239          *     ip's.
1240          */
1241         if (cdip == fdip &&
1242             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1243                 if (nlinks) {
1244                         hammer2_chain_modify(trans, &chain, 0);
1245                         chain->data->ipdata.nlinks += nlinks;
1246                 }
1247                 error = 0;
1248                 goto done;
1249         }
1250
1251         /*
1252          * We either have to move an existing hardlink target or we have
1253          * to create a fresh hardlink target.
1254          *
1255          * Hardlink targets are hidden inodes in a parent directory common
1256          * to all directory entries referencing the hardlink.
1257          */
1258         nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1259
1260         if (error == 0) {
1261                 /*
1262                  * Bump nlinks on duplicated hidden inode, repoint
1263                  * ip->chain.
1264                  */
1265                 hammer2_chain_modify(trans, &nchain, 0);
1266                 nchain->data->ipdata.nlinks += nlinks;
1267                 hammer2_inode_repoint(ip, cdip, nchain);
1268
1269                 /*
1270                  * If the old chain is not a hardlink target then replace
1271                  * it with a OBJTYPE_HARDLINK pointer.
1272                  *
1273                  * If the old chain IS a hardlink target then delete it.
1274                  */
1275                 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1276                         /*
1277                          * Replace original non-hardlink that's been dup'd
1278                          * with a special hardlink directory entry.  We must
1279                          * set the DIRECTDATA flag to prevent sub-chains
1280                          * from trying to synchronize to the inode if the
1281                          * file is extended afterwords.
1282                          */
1283                         hammer2_chain_modify(trans, &chain, 0);
1284                         ipdata = &chain->data->ipdata;
1285                         ipdata->target_type = ipdata->type;
1286                         ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1287                         ipdata->uflags = 0;
1288                         ipdata->rmajor = 0;
1289                         ipdata->rminor = 0;
1290                         ipdata->ctime = 0;
1291                         ipdata->mtime = 0;
1292                         ipdata->atime = 0;
1293                         ipdata->btime = 0;
1294                         bzero(&ipdata->uid, sizeof(ipdata->uid));
1295                         bzero(&ipdata->gid, sizeof(ipdata->gid));
1296                         ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1297                         ipdata->cap_flags = 0;
1298                         ipdata->mode = 0;
1299                         ipdata->size = 0;
1300                         ipdata->nlinks = 1;
1301                         ipdata->iparent = 0;    /* XXX */
1302                         ipdata->pfs_type = 0;
1303                         ipdata->pfs_inum = 0;
1304                         bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1305                         bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1306                         ipdata->data_quota = 0;
1307                         ipdata->data_count = 0;
1308                         ipdata->inode_quota = 0;
1309                         ipdata->inode_count = 0;
1310                         ipdata->attr_tid = 0;
1311                         ipdata->dirent_tid = 0;
1312                         bzero(&ipdata->u, sizeof(ipdata->u));
1313                         /* XXX transaction ids */
1314                 } else {
1315                         kprintf("DELETE INVISIBLE\n");
1316                         for (;;) {
1317                                 parent = chain->parent;
1318                                 hammer2_chain_ref(parent);
1319                                 hammer2_chain_ref(chain);
1320                                 hammer2_chain_unlock(chain);
1321                                 hammer2_chain_lock(parent,
1322                                                    HAMMER2_RESOLVE_ALWAYS);
1323                                 hammer2_chain_lock(chain,
1324                                                    HAMMER2_RESOLVE_ALWAYS);
1325                                 hammer2_chain_drop(chain);
1326                                 if (chain->parent == parent)
1327                                         break;
1328                                 hammer2_chain_unlock(parent);
1329                                 hammer2_chain_drop(parent);
1330                         }
1331                         hammer2_chain_delete(trans, parent, chain);
1332                         hammer2_chain_unlock(parent);
1333                         hammer2_chain_drop(parent);
1334                 }
1335
1336                 /*
1337                  * Return the new chain.
1338                  */
1339                 hammer2_chain_unlock(chain);
1340                 chain = nchain;
1341         } else {
1342                 /*
1343                  * Return an error
1344                  */
1345                 hammer2_chain_unlock(chain);
1346                 chain = NULL;
1347         }
1348
1349         /*
1350          * Cleanup, chain/nchain already dealt with.
1351          */
1352 done:
1353         *chainp = chain;
1354         hammer2_inode_drop(cdip);
1355
1356         return (error);
1357 }
1358
1359 /*
1360  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1361  * inode while (*chainp) points to the resolved (hidden hardlink
1362  * target) inode.  In this situation when nlinks is 1 we wish to
1363  * deconsolidate the hardlink, moving it back to the directory that now
1364  * represents the only remaining link.
1365  */
1366 int
1367 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1368                                hammer2_inode_t *dip,
1369                                hammer2_chain_t **chainp,
1370                                hammer2_chain_t **ochainp)
1371 {
1372         if (*ochainp == NULL)
1373                 return (0);
1374         /* XXX */
1375         return (0);
1376 }
1377
1378 /*
1379  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1380  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1381  * the *chainp and return a new locked *chainp representing the file target
1382  * (the original *chainp will be unlocked).
1383  *
1384  * When a match is found the chain representing the original HARDLINK
1385  * will be returned in *ochainp with a ref, but not locked.
1386  *
1387  * When no match is found *chainp is set to NULL and EIO is returned.
1388  * (*ochainp) will still be set to the original chain with a ref but not
1389  * locked.
1390  */
1391 int
1392 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1393                       hammer2_chain_t **ochainp)
1394 {
1395         hammer2_chain_t *chain = *chainp;
1396         hammer2_chain_t *parent;
1397         hammer2_inode_t *ip;
1398         hammer2_inode_t *pip;
1399         hammer2_key_t lhc;
1400
1401         pip = dip;
1402         hammer2_inode_ref(pip);         /* for loop */
1403         hammer2_chain_ref(chain);       /* for (*ochainp) */
1404         *ochainp = chain;
1405
1406         /*
1407          * Locate the hardlink.  pip is referenced and not locked,
1408          * ipp.
1409          *
1410          * chain is reused.
1411          */
1412         lhc = chain->data->ipdata.inum;
1413         hammer2_chain_unlock(chain);
1414         chain = NULL;
1415
1416         while ((ip = pip) != NULL) {
1417                 parent = hammer2_inode_lock_ex(ip);
1418                 hammer2_inode_drop(ip);                 /* loop */
1419                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1420                 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1421                 hammer2_chain_lookup_done(parent);      /* discard parent */
1422                 if (chain)
1423                         break;
1424                 pip = ip->pip;          /* safe, ip held locked */
1425                 if (pip)
1426                         hammer2_inode_ref(pip);         /* loop */
1427                 hammer2_inode_unlock_ex(ip, NULL);
1428         }
1429
1430         /*
1431          * chain is locked, ip is locked.  Unlock ip, return the locked
1432          * chain.  *ipp is already set w/a ref count and not locked.
1433          *
1434          * (parent is already unlocked).
1435          */
1436         if (ip)
1437                 hammer2_inode_unlock_ex(ip, NULL);
1438         *chainp = chain;
1439         if (chain) {
1440                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1441                 /* already locked */
1442                 return (0);
1443         } else {
1444                 return (EIO);
1445         }
1446 }
1447
1448 /*
1449  * Find the directory common to both fdip and tdip, hold and return
1450  * its inode.
1451  */
1452 hammer2_inode_t *
1453 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1454 {
1455         hammer2_inode_t *scan1;
1456         hammer2_inode_t *scan2;
1457
1458         /*
1459          * We used to have a depth field but it complicated matters too
1460          * much for directory renames.  So now its ugly.  Check for
1461          * simple cases before giving up and doing it the expensive way.
1462          *
1463          * XXX need a bottom-up topology stability lock
1464          */
1465         if (fdip == tdip || fdip == tdip->pip) {
1466                 hammer2_inode_ref(fdip);
1467                 return(fdip);
1468         }
1469         if (fdip->pip == tdip) {
1470                 hammer2_inode_ref(tdip);
1471                 return(tdip);
1472         }
1473
1474         /*
1475          * XXX not MPSAFE
1476          */
1477         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1478                 scan2 = tdip;
1479                 while (scan2->pmp == tdip->pmp) {
1480                         if (scan1 == scan2) {
1481                                 hammer2_inode_ref(scan1);
1482                                 return(scan1);
1483                         }
1484                         scan2 = scan2->pip;
1485                         if (scan2 == NULL)
1486                                 break;
1487                 }
1488         }
1489         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1490               fdip, tdip);
1491         /* NOT REACHED */
1492         return(NULL);
1493 }