sys/vfs/hammer2/hammer2_inode.c

   1 /*
   2  * Copyright (c) 2011-2012 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 /*
  45  * Adding a ref to an inode is only legal if the inode already has at least
  46  * one ref.
  47  */
  48 void
  49 hammer2_inode_ref(hammer2_inode_t *ip)
  50 {
  51         hammer2_chain_ref(ip->hmp, &ip->chain);
  52 }
  53
  54 /*
  55  * Drop an inode reference, freeing the inode when the last reference goes
  56  * away.
  57  */
  58 void
  59 hammer2_inode_drop(hammer2_inode_t *ip)
  60 {
  61         hammer2_chain_drop(ip->hmp, &ip->chain);
  62 }
  63
  64 /*
  65  * Get the vnode associated with the given inode, allocating the vnode if
  66  * necessary.  The vnode will be returned exclusively locked.
  67  *
  68  * The caller must lock the inode (shared or exclusive).
  69  *
  70  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
  71  * races.
  72  */
  73 struct vnode *
  74 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
  75 {
  76         struct vnode *vp;
  77         hammer2_pfsmount_t *pmp;
  78         ccms_state_t ostate;
  79
  80         pmp = ip->pmp;
  81         KKASSERT(pmp != NULL);
  82         *errorp = 0;
  83
  84         for (;;) {
  85                 /*
  86                  * Attempt to reuse an existing vnode assignment.  It is
  87                  * possible to race a reclaim so the vget() may fail.  The
  88                  * inode must be unlocked during the vget() to avoid a
  89                  * deadlock against a reclaim.
  90                  */
  91                 vp = ip->vp;
  92                 if (vp) {
  93                         /*
  94                          * Inode must be unlocked during the vget() to avoid
  95                          * possible deadlocks, vnode is held to prevent
  96                          * destruction during the vget().  The vget() can
  97                          * still fail if we lost a reclaim race on the vnode.
  98                          */
  99                         vhold_interlocked(vp);
 100                         ccms_thread_unlock(&ip->chain.cst);
 101                         if (vget(vp, LK_EXCLUSIVE)) {
 102                                 vdrop(vp);
 103                                 ccms_thread_lock(&ip->chain.cst,
 104                                                  CCMS_STATE_EXCLUSIVE);
 105                                 continue;
 106                         }
 107                         ccms_thread_lock(&ip->chain.cst, CCMS_STATE_EXCLUSIVE);
 108                         vdrop(vp);
 109                         /* vp still locked and ref from vget */
 110                         *errorp = 0;
 111                         break;
 112                 }
 113
 114                 /*
 115                  * No vnode exists, allocate a new vnode.  Beware of
 116                  * allocation races.  This function will return an
 117                  * exclusively locked and referenced vnode.
 118                  */
 119                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
 120                 if (*errorp) {
 121                         vp = NULL;
 122                         break;
 123                 }
 124
 125                 /*
 126                  * Lock the inode and check for an allocation race.
 127                  */
 128                 ostate = ccms_thread_lock_upgrade(&ip->chain.cst);
 129                 if (ip->vp != NULL) {
 130                         vp->v_type = VBAD;
 131                         vx_put(vp);
 132                         ccms_thread_lock_restore(&ip->chain.cst, ostate);
 133                         continue;
 134                 }
 135
 136                 switch (ip->ip_data.type) {
 137                 case HAMMER2_OBJTYPE_DIRECTORY:
 138                         vp->v_type = VDIR;
 139                         break;
 140                 case HAMMER2_OBJTYPE_REGFILE:
 141                         vp->v_type = VREG;
 142                         vinitvmio(vp, ip->ip_data.size,
 143                                   HAMMER2_LBUFSIZE,
 144                                   (int)ip->ip_data.size & HAMMER2_LBUFMASK);
 145                         break;
 146                 case HAMMER2_OBJTYPE_SOFTLINK:
 147                         /*
 148                          * XXX for now we are using the generic file_read
 149                          * and file_write code so we need a buffer cache
 150                          * association.
 151                          */
 152                         vp->v_type = VLNK;
 153                         vinitvmio(vp, ip->ip_data.size,
 154                                   HAMMER2_LBUFSIZE,
 155                                   (int)ip->ip_data.size & HAMMER2_LBUFMASK);
 156                         break;
 157                 /* XXX FIFO */
 158                 default:
 159                         panic("hammer2: unhandled objtype %d",
 160                               ip->ip_data.type);
 161                         break;
 162                 }
 163
 164                 if (ip == pmp->iroot)
 165                         vsetflags(vp, VROOT);
 166
 167                 vp->v_data = ip;
 168                 ip->vp = vp;
 169                 hammer2_chain_ref(ip->hmp, &ip->chain); /* vp association */
 170                 ccms_thread_lock_restore(&ip->chain.cst, ostate);
 171                 break;
 172         }
 173
 174         /*
 175          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
 176          */
 177         if (hammer2_debug & 0x0002) {
 178                 kprintf("igetv vp %p refs %d aux %d\n",
 179                         vp, vp->v_sysref.refcnt, vp->v_auxrefs);
 180         }
 181         return (vp);
 182 }
 183
 184 /*
 185  * Create a new inode in the specified directory using the vattr to
 186  * figure out the type of inode.
 187  *
 188  * If no error occurs the new inode with its chain locked is returned in
 189  * *nipp, otherwise an error is returned and *nipp is set to NULL.
 190  *
 191  * If vap and/or cred are NULL the related fields are not set and the
 192  * inode type defaults to a directory.  This is used when creating PFSs
 193  * under the super-root, so the inode number is set to 1 in this case.
 194  */
 195 int
 196 hammer2_inode_create(hammer2_inode_t *dip,
 197                      struct vattr *vap, struct ucred *cred,
 198                      const uint8_t *name, size_t name_len,
 199                      hammer2_inode_t **nipp)
 200 {
 201         hammer2_mount_t *hmp = dip->hmp;
 202         hammer2_chain_t *chain;
 203         hammer2_chain_t *parent;
 204         hammer2_inode_t *nip;
 205         hammer2_key_t lhc;
 206         int error;
 207         uid_t xuid;
 208
 209         lhc = hammer2_dirhash(name, name_len);
 210
 211         /*
 212          * Locate the inode or indirect block to create the new
 213          * entry in.  At the same time check for key collisions
 214          * and iterate until we don't get one.
 215          */
 216         parent = &dip->chain;
 217         hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 218
 219         error = 0;
 220         while (error == 0) {
 221                 chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
 222                 if (chain == NULL)
 223                         break;
 224                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
 225                         error = ENOSPC;
 226                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 227                         error = ENOSPC;
 228                 hammer2_chain_unlock(hmp, chain);
 229                 chain = NULL;
 230                 ++lhc;
 231         }
 232         if (error == 0) {
 233                 chain = hammer2_chain_create(hmp, parent, NULL, lhc, 0,
 234                                              HAMMER2_BREF_TYPE_INODE,
 235                                              HAMMER2_INODE_BYTES);
 236                 if (chain == NULL)
 237                         error = EIO;
 238         }
 239         hammer2_chain_unlock(hmp, parent);
 240
 241         /*
 242          * Handle the error case
 243          */
 244         if (error) {
 245                 KKASSERT(chain == NULL);
 246                 *nipp = NULL;
 247                 return (error);
 248         }
 249
 250         /*
 251          * Set up the new inode
 252          */
 253         nip = chain->u.ip;
 254         *nipp = nip;
 255
 256         hammer2_voldata_lock(hmp);
 257         if (vap) {
 258                 nip->ip_data.type = hammer2_get_obj_type(vap->va_type);
 259                 nip->ip_data.inum = hmp->voldata.alloc_tid++;
 260                 /* XXX modify/lock */
 261         } else {
 262                 nip->ip_data.type = HAMMER2_OBJTYPE_DIRECTORY;
 263                 nip->ip_data.inum = 1;
 264         }
 265         hammer2_voldata_unlock(hmp);
 266         nip->ip_data.version = HAMMER2_INODE_VERSION_ONE;
 267         hammer2_update_time(&nip->ip_data.ctime);
 268         nip->ip_data.mtime = nip->ip_data.ctime;
 269         if (vap)
 270                 nip->ip_data.mode = vap->va_mode;
 271         nip->ip_data.nlinks = 1;
 272         if (vap) {
 273                 if (dip) {
 274                         xuid = hammer2_to_unix_xid(&dip->ip_data.uid);
 275                         xuid = vop_helper_create_uid(dip->pmp->mp,
 276                                                      dip->ip_data.mode,
 277                                                      xuid,
 278                                                      cred,
 279                                                      &vap->va_mode);
 280                 } else {
 281                         xuid = 0;
 282                 }
 283                 if (vap->va_vaflags & VA_UID_UUID_VALID)
 284                         nip->ip_data.uid = vap->va_uid_uuid;
 285                 else if (vap->va_uid != (uid_t)VNOVAL)
 286                         hammer2_guid_to_uuid(&nip->ip_data.uid, vap->va_uid);
 287                 else
 288                         hammer2_guid_to_uuid(&nip->ip_data.uid, xuid);
 289
 290                 if (vap->va_vaflags & VA_GID_UUID_VALID)
 291                         nip->ip_data.gid = vap->va_gid_uuid;
 292                 else if (vap->va_gid != (gid_t)VNOVAL)
 293                         hammer2_guid_to_uuid(&nip->ip_data.gid, vap->va_gid);
 294                 else if (dip)
 295                         nip->ip_data.gid = dip->ip_data.gid;
 296         }
 297
 298         /*
 299          * Regular files and softlinks allow a small amount of data to be
 300          * directly embedded in the inode.  This flag will be cleared if
 301          * the size is extended past the embedded limit.
 302          */
 303         if (nip->ip_data.type == HAMMER2_OBJTYPE_REGFILE ||
 304             nip->ip_data.type == HAMMER2_OBJTYPE_SOFTLINK) {
 305                 nip->ip_data.op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
 306         }
 307
 308         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 309         bcopy(name, nip->ip_data.filename, name_len);
 310         nip->ip_data.name_key = lhc;
 311         nip->ip_data.name_len = name_len;
 312
 313         return (0);
 314 }
 315
 316 /*
 317  * Duplicate the specified existing inode in the specified target directory.
 318  * If name is NULL the inode is duplicated as a hidden directory entry.
 319  *
 320  * Returns the new inode.  The old inode is left alone.
 321  *
 322  * XXX name needs to be NULL for now.
 323  */
 324 int
 325 hammer2_inode_duplicate(hammer2_inode_t *dip, hammer2_inode_t *oip,
 326                         hammer2_inode_t **nipp,
 327                         const uint8_t *name, size_t name_len)
 328 {
 329         hammer2_mount_t *hmp = dip->hmp;
 330         hammer2_inode_t *nip;
 331         hammer2_chain_t *parent;
 332         hammer2_chain_t *chain;
 333         hammer2_key_t lhc;
 334         int error;
 335
 336         if (name) {
 337                 lhc = hammer2_dirhash(name, name_len);
 338         } else {
 339                 lhc = oip->ip_data.inum;
 340                 KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
 341         }
 342
 343         /*
 344          * Locate the inode or indirect block to create the new
 345          * entry in.  At the same time check for key collisions
 346          * and iterate until we don't get one.
 347          */
 348         nip = NULL;
 349         parent = &dip->chain;
 350         hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 351
 352         error = 0;
 353         while (error == 0) {
 354                 chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
 355                 if (chain == NULL)
 356                         break;
 357                 /* XXX bcmp name if not NULL */
 358                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 359                         error = ENOSPC;
 360                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0) /* shouldn't happen */
 361                         error = ENOSPC;
 362                 hammer2_chain_unlock(hmp, chain);
 363                 chain = NULL;
 364                 ++lhc;
 365         }
 366
 367         /*
 368          * Create entry in common parent directory.
 369          */
 370         if (error == 0) {
 371                 chain = hammer2_chain_create(hmp, parent, NULL, lhc, 0,
 372                                              HAMMER2_BREF_TYPE_INODE /* n/a */,
 373                                              HAMMER2_INODE_BYTES);   /* n/a */
 374                 if (chain == NULL)
 375                         error = EIO;
 376         }
 377         hammer2_chain_unlock(hmp, parent);
 378
 379         /*
 380          * Handle the error case
 381          */
 382         if (error) {
 383                 KKASSERT(chain == NULL);
 384                 return (error);
 385         }
 386
 387         /*
 388          * XXX This is currently a horrible hack.  Well, if we wanted to
 389          *     duplicate a file, i.e. as in a snapshot, we definitely
 390          *     would have to flush it first.
 391          *
 392          *     For hardlink target generation we can theoretically move any
 393          *     active chain structures without flushing, but that gets really
 394          *     iffy for code which follows chain->parent and ip->pip links.
 395          *
 396          * XXX only works with files.  Duplicating a directory hierarchy
 397          *     requires a flush but doesn't deal with races post-flush.
 398          *     Well, it would work I guess, but you might catch some files
 399          *     mid-operation.
 400          *
 401          * We cannot leave oip with any in-memory chains because (for a
 402          * hardlink), oip will become a OBJTYPE_HARDLINK which is just a
 403          * pointer to the real hardlink's inum and can't have any sub-chains.
 404          * XXX might be 0-ref chains left.
 405          */
 406         hammer2_inode_lock_ex(oip);
 407         hammer2_chain_flush(hmp, &oip->chain, 0);
 408         hammer2_inode_unlock_ex(oip);
 409         /*KKASSERT(RB_EMPTY(&oip->chain.rbhead));*/
 410
 411         nip = chain->u.ip;
 412         hammer2_chain_modify(hmp, chain, 0);
 413         nip->ip_data = oip->ip_data;    /* sync media data after flush */
 414
 415         if (name) {
 416                 /*
 417                  * Directory entries are inodes so if the name has changed
 418                  * we have to update the inode.
 419                  */
 420                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 421                 bcopy(name, nip->ip_data.filename, name_len);
 422                 nip->ip_data.name_key = lhc;
 423                 nip->ip_data.name_len = name_len;
 424         } else {
 425                 /*
 426                  * Directory entries are inodes but this is a hidden hardlink
 427                  * target.  The name isn't used but to ease debugging give it
 428                  * a name after its inode number.
 429                  */
 430                 ksnprintf(nip->ip_data.filename, sizeof(nip->ip_data.filename),
 431                           "0x%016jx", (intmax_t)nip->ip_data.inum);
 432                 nip->ip_data.name_len = strlen(nip->ip_data.filename);
 433                 nip->ip_data.name_key = lhc;
 434         }
 435         *nipp = nip;
 436
 437         return (0);
 438 }
 439
 440
 441 /*
 442  * Connect inode (oip) to the specified directory using the specified name.
 443  * (oip) must be locked.
 444  *
 445  * If (oip) is not currently connected we simply connect it up.
 446  *
 447  * If (oip) is already connected we create a OBJTYPE_HARDLINK entry which
 448  * points to (oip)'s inode number.  (oip) is expected to be the terminus of
 449  * the hardlink sitting as a hidden file in a common parent directory
 450  * in this situation.
 451  */
 452 int
 453 hammer2_inode_connect(hammer2_inode_t *dip, hammer2_inode_t *oip,
 454                       const uint8_t *name, size_t name_len)
 455 {
 456         hammer2_mount_t *hmp = dip->hmp;
 457         hammer2_chain_t *chain;
 458         hammer2_chain_t *parent;
 459         hammer2_inode_t *nip;
 460         hammer2_key_t lhc;
 461         int error;
 462         int hlink;
 463
 464         /*
 465          * (oip) is the terminus of the hardlink sitting in the common
 466          * parent directory.  This means that if oip->pip != dip then
 467          * the already locked oip is ABOVE dip.
 468          *
 469          * But if the common parent directory IS dip, then we would have
 470          * a lock-order reversal and must rearrange the lock ordering.
 471          * For now the caller deals with this for us by locking dip in
 472          * that case (and our lock here winds up just being recursive)
 473          */
 474         parent = &dip->chain;
 475         if (oip->pip == dip) {
 476                 hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 477                 hammer2_chain_lock(hmp, &oip->chain, HAMMER2_RESOLVE_ALWAYS);
 478         } else {
 479                 hammer2_chain_lock(hmp, &oip->chain, HAMMER2_RESOLVE_ALWAYS);
 480                 hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 481         }
 482
 483
 484         lhc = hammer2_dirhash(name, name_len);
 485         hlink = (oip->chain.parent != NULL);
 486
 487         /*
 488          * In fake mode flush oip so we can just snapshot it downbelow.
 489          */
 490         if (hlink && hammer2_hardlink_enable < 0)
 491                 hammer2_chain_flush(hmp, &oip->chain, 0);
 492
 493         /*
 494          * Locate the inode or indirect block to create the new
 495          * entry in.  At the same time check for key collisions
 496          * and iterate until we don't get one.
 497          */
 498         error = 0;
 499         while (error == 0) {
 500                 chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
 501                 if (chain == NULL)
 502                         break;
 503                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 504                         error = ENOSPC;
 505                 hammer2_chain_unlock(hmp, chain);
 506                 chain = NULL;
 507                 ++lhc;
 508         }
 509
 510         /*
 511          * Passing a non-NULL chain to hammer2_chain_create() reconnects the
 512          * existing chain instead of creating a new one.  The chain's bref
 513          * will be properly updated.
 514          */
 515         if (error == 0) {
 516                 if (hlink) {
 517                         chain = hammer2_chain_create(hmp, parent,
 518                                                      NULL, lhc, 0,
 519                                                      HAMMER2_BREF_TYPE_INODE,
 520                                                      HAMMER2_INODE_BYTES);
 521                 } else {
 522                         chain = hammer2_chain_create(hmp, parent,
 523                                                      &oip->chain, lhc, 0,
 524                                                      HAMMER2_BREF_TYPE_INODE,
 525                                                      HAMMER2_INODE_BYTES);
 526                         if (chain)
 527                                 KKASSERT(chain == &oip->chain);
 528                 }
 529                 if (chain == NULL)
 530                         error = EIO;
 531         }
 532         hammer2_chain_unlock(hmp, parent);
 533
 534         /*
 535          * Handle the error case
 536          */
 537         if (error) {
 538                 KKASSERT(chain == NULL);
 539                 hammer2_chain_unlock(hmp, &oip->chain);
 540                 return (error);
 541         }
 542
 543         /*
 544          * Directory entries are inodes so if the name has changed we have
 545          * to update the inode.
 546          *
 547          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
 548          * chain, the caller will access the hardlink via the actual hardlink
 549          * target file and not the hardlink pointer entry.
 550          */
 551         if (hlink && hammer2_hardlink_enable >= 0) {
 552                 /*
 553                  * Create the HARDLINK pointer.  oip represents the hardlink
 554                  * target in this situation.
 555                  */
 556                 nip = chain->u.ip;
 557                 hammer2_chain_modify(hmp, chain, 0);
 558                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 559                 bcopy(name, nip->ip_data.filename, name_len);
 560                 nip->ip_data.name_key = lhc;
 561                 nip->ip_data.name_len = name_len;
 562                 nip->ip_data.target_type = oip->ip_data.type;
 563                 nip->ip_data.type = HAMMER2_OBJTYPE_HARDLINK;
 564                 nip->ip_data.inum = oip->ip_data.inum;
 565                 nip->ip_data.nlinks = 1;
 566                 kprintf("created hardlink %*.*s\n",
 567                         (int)name_len, (int)name_len, name);
 568                 hammer2_chain_unlock(hmp, chain);
 569         } else if (hlink && hammer2_hardlink_enable < 0) {
 570                 /*
 571                  * Create a snapshot (hardlink fake mode for debugging).
 572                  */
 573                 nip = chain->u.ip;
 574                 nip->ip_data = oip->ip_data;
 575                 hammer2_chain_modify(hmp, chain, 0);
 576                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 577                 bcopy(name, nip->ip_data.filename, name_len);
 578                 nip->ip_data.name_key = lhc;
 579                 nip->ip_data.name_len = name_len;
 580                 kprintf("created fake hardlink %*.*s\n",
 581                         (int)name_len, (int)name_len, name);
 582                 hammer2_chain_unlock(hmp, chain);
 583         } else {
 584                 /*
 585                  * Normally disconnected inode (e.g. during a rename) that
 586                  * was reconnected.  We must fixup the name stored in
 587                  * oip.
 588                  *
 589                  * We are using oip as chain, already locked by caller,
 590                  * do not unlock it.
 591                  */
 592                 hammer2_chain_modify(hmp, chain, 0);
 593                 if (oip->ip_data.name_len != name_len ||
 594                     bcmp(oip->ip_data.filename, name, name_len) != 0) {
 595                         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 596                         bcopy(name, oip->ip_data.filename, name_len);
 597                         oip->ip_data.name_key = lhc;
 598                         oip->ip_data.name_len = name_len;
 599                 }
 600                 oip->ip_data.nlinks = 1;
 601         }
 602         hammer2_chain_unlock(hmp, &oip->chain);
 603         return (0);
 604 }
 605
 606 /*
 607  * Unlink the file from the specified directory inode.  The directory inode
 608  * does not need to be locked.
 609  *
 610  * isdir determines whether a directory/non-directory check should be made.
 611  * No check is made if isdir is set to -1.
 612  */
 613 int
 614 hammer2_unlink_file(hammer2_inode_t *dip,
 615                     const uint8_t *name, size_t name_len,
 616                     int isdir, hammer2_inode_t *retain_ip)
 617 {
 618         hammer2_mount_t *hmp;
 619         hammer2_chain_t *parent;
 620         hammer2_chain_t *chain;
 621         hammer2_chain_t *dparent;
 622         hammer2_chain_t *dchain;
 623         hammer2_key_t lhc;
 624         hammer2_inode_t *ip;
 625         hammer2_inode_t *oip;
 626         int error;
 627         uint8_t type;
 628
 629         error = 0;
 630         oip = NULL;
 631         hmp = dip->hmp;
 632         lhc = hammer2_dirhash(name, name_len);
 633
 634         /*
 635          * Search for the filename in the directory
 636          */
 637         parent = &dip->chain;
 638         hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 639         chain = hammer2_chain_lookup(hmp, &parent,
 640                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
 641                                      HAMMER2_LOOKUP_MAYDELETE);
 642         while (chain) {
 643                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
 644                     chain->u.ip &&
 645                     name_len == chain->data->ipdata.name_len &&
 646                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
 647                         break;
 648                 }
 649                 chain = hammer2_chain_next(hmp, &parent, chain,
 650                                            lhc, lhc + HAMMER2_DIRHASH_LOMASK,
 651                                            HAMMER2_LOOKUP_MAYDELETE);
 652         }
 653
 654         /*
 655          * Not found or wrong type (isdir < 0 disables the type check).
 656          */
 657         if (chain == NULL) {
 658                 hammer2_chain_unlock(hmp, parent);
 659                 return ENOENT;
 660         }
 661         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK)
 662                 type = chain->data->ipdata.target_type;
 663
 664         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
 665                 error = ENOTDIR;
 666                 goto done;
 667         }
 668         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
 669                 error = EISDIR;
 670                 goto done;
 671         }
 672
 673         /*
 674          * Hardlink must be resolved.  We can't hold parent locked while we
 675          * do this or we could deadlock.
 676          */
 677         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
 678                 hammer2_chain_unlock(hmp, parent);
 679                 parent = NULL;
 680                 error = hammer2_hardlink_find(dip, &chain, &oip);
 681         }
 682
 683         /*
 684          * If this is a directory the directory must be empty.  However, if
 685          * isdir < 0 we are doing a rename and the directory does not have
 686          * to be empty.
 687          *
 688          * NOTE: We check the full key range here which covers both visible
 689          *       and invisible entries.  Theoretically there should be no
 690          *       invisible (hardlink target) entries if there are no visible
 691          *       entries.
 692          */
 693         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir >= 0) {
 694                 dparent = chain;
 695                 hammer2_chain_lock(hmp, dparent, HAMMER2_RESOLVE_ALWAYS);
 696                 dchain = hammer2_chain_lookup(hmp, &dparent,
 697                                               0, (hammer2_key_t)-1,
 698                                               HAMMER2_LOOKUP_NODATA);
 699                 if (dchain) {
 700                         hammer2_chain_unlock(hmp, dchain);
 701                         hammer2_chain_unlock(hmp, dparent);
 702                         error = ENOTEMPTY;
 703                         goto done;
 704                 }
 705                 hammer2_chain_unlock(hmp, dparent);
 706                 dparent = NULL;
 707                 /* dchain NULL */
 708         }
 709
 710         /*
 711          * Ok, we can now unlink the chain.  We always decrement nlinks even
 712          * if the entry can be deleted in case someone has the file open and
 713          * does an fstat().
 714          *
 715          * The chain itself will no longer be in the on-media topology but
 716          * can still be flushed to the media (e.g. if an open descriptor
 717          * remains).  When the last vnode/ip ref goes away the chain will
 718          * be marked unmodified, avoiding any further (now unnecesary) I/O.
 719          */
 720         if (oip) {
 721                 /*
 722                  * If this was a hardlink we first delete the hardlink
 723                  * pointer entry.
 724                  */
 725                 parent = oip->chain.parent;
 726                 hammer2_chain_lock_pair(hmp, parent, &oip->chain,
 727                                         HAMMER2_RESOLVE_ALWAYS |
 728                                         HAMMER2_RESOLVE_MAYDELETE);
 729                 hammer2_chain_delete(hmp, parent, &oip->chain,
 730                                     (retain_ip == oip));
 731                 hammer2_chain_unlock(hmp, &oip->chain);
 732                 hammer2_chain_unlock(hmp, parent);
 733                 parent = NULL;
 734
 735                 /*
 736                  * Then decrement nlinks on hardlink target.
 737                  */
 738                 ip = chain->u.ip;
 739                 if (ip->ip_data.nlinks == 1) {
 740                         dparent = chain->parent;
 741                         hammer2_chain_ref(hmp, chain);
 742                         hammer2_chain_unlock(hmp, chain);
 743                         hammer2_chain_lock_pair(hmp, dparent, chain,
 744                                            HAMMER2_RESOLVE_ALWAYS |
 745                                            HAMMER2_RESOLVE_MAYDELETE);
 746                         hammer2_chain_drop(hmp, chain);
 747                         hammer2_chain_modify(hmp, chain, 0);
 748                         --ip->ip_data.nlinks;
 749                         hammer2_chain_delete(hmp, dparent, chain, 0);
 750                         hammer2_chain_unlock(hmp, dparent);
 751                 } else {
 752                         hammer2_chain_modify(hmp, chain, 0);
 753                         --ip->ip_data.nlinks;
 754                 }
 755         } else {
 756                 /*
 757                  * Otherwise this was not a hardlink and we can just
 758                  * remove the entry and decrement nlinks.
 759                  */
 760                 ip = chain->u.ip;
 761                 hammer2_chain_modify(hmp, chain, 0);
 762                 --ip->ip_data.nlinks;
 763                 hammer2_chain_delete(hmp, parent, chain,
 764                                      (retain_ip == ip));
 765         }
 766
 767         error = 0;
 768
 769 done:
 770         if (chain)
 771                 hammer2_chain_unlock(hmp, chain);
 772         if (parent)
 773                 hammer2_chain_unlock(hmp, parent);
 774         if (oip)
 775                 hammer2_chain_drop(oip->hmp, &oip->chain);
 776
 777         return error;
 778 }
 779
 780 /*
 781  * Calculate the allocation size for the file fragment straddling EOF
 782  */
 783 int
 784 hammer2_inode_calc_alloc(hammer2_key_t filesize)
 785 {
 786         int frag = (int)filesize & HAMMER2_PBUFMASK;
 787         int radix;
 788
 789         if (frag == 0)
 790                 return(0);
 791         for (radix = HAMMER2_MINALLOCRADIX; frag > (1 << radix); ++radix)
 792                 ;
 793         return (radix);
 794 }
 795
 796 void
 797 hammer2_inode_lock_nlinks(hammer2_inode_t *ip)
 798 {
 799         hammer2_chain_ref(ip->hmp, &ip->chain);
 800 }
 801
 802 void
 803 hammer2_inode_unlock_nlinks(hammer2_inode_t *ip)
 804 {
 805         hammer2_chain_drop(ip->hmp, &ip->chain);
 806 }
 807
 808 /*
 809  * Consolidate for hard link creation.  This moves the specified terminal
 810  * hardlink inode to a directory common to its current directory and tdip
 811  * if necessary, replacing *ipp with the new inode chain element and
 812  * modifying the original inode chain element to OBJTYPE_HARDLINK.
 813  *
 814  * If the original inode chain element was a prior incarnation of a hidden
 815  * inode it can simply be deleted instead of converted.
 816  *
 817  * (*ipp)'s nlinks field is locked on entry and the new (*ipp)'s nlinks
 818  * field will be locked on return (with the original's unlocked).
 819  *
 820  * The link count is bumped if requested.
 821  */
 822 int
 823 hammer2_hardlink_consolidate(hammer2_inode_t **ipp, hammer2_inode_t *tdip)
 824 {
 825         hammer2_mount_t *hmp;
 826         hammer2_inode_t *oip = *ipp;
 827         hammer2_inode_t *nip = NULL;
 828         hammer2_inode_t *fdip;
 829         hammer2_inode_t *cdip;
 830         hammer2_chain_t *parent;
 831         int error;
 832
 833         hmp = tdip->hmp;
 834
 835         if (hammer2_hardlink_enable < 0)
 836                 return (0);
 837         if (hammer2_hardlink_enable == 0)
 838                 return (ENOTSUP);
 839
 840         fdip = oip->pip;
 841         cdip = hammer2_inode_common_parent(hmp, fdip, tdip);
 842
 843         /*
 844          * Nothing to do (except bump the link count) if the hardlink has
 845          * already been consolidated in the correct place.
 846          */
 847         if (cdip == fdip &&
 848             (oip->ip_data.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
 849                 kprintf("hardlink already consolidated correctly\n");
 850                 nip = oip;
 851                 hammer2_inode_lock_ex(nip);
 852                 hammer2_chain_modify(hmp, &nip->chain, 0);
 853                 ++nip->ip_data.nlinks;
 854                 hammer2_inode_unlock_ex(nip);
 855                 hammer2_inode_drop(cdip);
 856                 return (0);
 857         }
 858
 859         /*
 860          * Create a hidden inode directory entry in the parent, copying
 861          * (*oip)'s state.  Then replace oip with OBJTYPE_HARDLINK.
 862          *
 863          * The duplication function will either flush or move any chains
 864          * under oip to the new hardlink target inode, retiring all chains
 865          * related to oip before returning.  XXX vp->ip races.
 866          */
 867         error = hammer2_inode_duplicate(cdip, oip, &nip, NULL, 0);
 868         if (error == 0) {
 869                 /*
 870                  * Bump nlinks on duplicated hidden inode.
 871                  */
 872                 kprintf("hardlink consolidation success in parent dir %s\n",
 873                         cdip->ip_data.filename);
 874                 hammer2_inode_lock_nlinks(nip);
 875                 hammer2_inode_unlock_nlinks(oip);
 876                 hammer2_chain_modify(hmp, &nip->chain, 0);
 877                 ++nip->ip_data.nlinks;
 878                 hammer2_inode_unlock_ex(nip);
 879
 880                 if (oip->ip_data.name_key & HAMMER2_DIRHASH_VISIBLE) {
 881                         /*
 882                          * Replace the old inode with an OBJTYPE_HARDLINK
 883                          * pointer.
 884                          */
 885                         hammer2_inode_lock_ex(oip);
 886                         hammer2_chain_modify(hmp, &oip->chain, 0);
 887                         oip->ip_data.target_type = oip->ip_data.type;
 888                         oip->ip_data.type = HAMMER2_OBJTYPE_HARDLINK;
 889                         oip->ip_data.uflags = 0;
 890                         oip->ip_data.rmajor = 0;
 891                         oip->ip_data.rminor = 0;
 892                         oip->ip_data.ctime = 0;
 893                         oip->ip_data.mtime = 0;
 894                         oip->ip_data.atime = 0;
 895                         oip->ip_data.btime = 0;
 896                         bzero(&oip->ip_data.uid, sizeof(oip->ip_data.uid));
 897                         bzero(&oip->ip_data.gid, sizeof(oip->ip_data.gid));
 898                         oip->ip_data.op_flags = HAMMER2_OPFLAG_DIRECTDATA;
 899                         oip->ip_data.cap_flags = 0;
 900                         oip->ip_data.mode = 0;
 901                         oip->ip_data.size = 0;
 902                         oip->ip_data.nlinks = 1;
 903                         oip->ip_data.iparent = 0;       /* XXX */
 904                         oip->ip_data.pfs_type = 0;
 905                         oip->ip_data.pfs_inum = 0;
 906                         bzero(&oip->ip_data.pfs_clid,
 907                               sizeof(oip->ip_data.pfs_clid));
 908                         bzero(&oip->ip_data.pfs_fsid,
 909                               sizeof(oip->ip_data.pfs_fsid));
 910                         oip->ip_data.data_quota = 0;
 911                         oip->ip_data.data_count = 0;
 912                         oip->ip_data.inode_quota = 0;
 913                         oip->ip_data.inode_count = 0;
 914                         oip->ip_data.attr_tid = 0;
 915                         oip->ip_data.dirent_tid = 0;
 916                         bzero(&oip->ip_data.u, sizeof(oip->ip_data.u));
 917                         /* XXX transaction ids */
 918
 919                         hammer2_inode_unlock_ex(oip);
 920                 } else {
 921                         /*
 922                          * The old inode was a hardlink target, which we
 923                          * have now moved.  We must delete it so the new
 924                          * hardlink target at a higher directory level
 925                          * becomes the only hardlink target for this inode.
 926                          */
 927                         kprintf("DELETE INVISIBLE\n");
 928                         parent = oip->chain.parent;
 929                         hammer2_chain_lock(hmp, parent,
 930                                            HAMMER2_RESOLVE_ALWAYS);
 931                         hammer2_chain_lock(hmp, &oip->chain,
 932                                            HAMMER2_RESOLVE_ALWAYS);
 933                         hammer2_chain_delete(hmp, parent, &oip->chain, 0);
 934                         hammer2_chain_unlock(hmp, &oip->chain);
 935                         hammer2_chain_unlock(hmp, parent);
 936                 }
 937                 *ipp = nip;
 938         } else {
 939                 KKASSERT(nip == NULL);
 940         }
 941         hammer2_inode_drop(cdip);
 942
 943         return (error);
 944 }
 945
 946 /*
 947  * If (*ipp) is non-NULL it points to the forward OBJTYPE_HARDLINK inode while
 948  * (*chainp) points to the resolved (hidden hardlink target) inode.  In this
 949  * situation when nlinks is 1 we wish to deconsolidate the hardlink, moving
 950  * it back to the directory that now represents the only remaining link.
 951  */
 952 int
 953 hammer2_hardlink_deconsolidate(hammer2_inode_t *dip, hammer2_chain_t **chainp,
 954                                hammer2_inode_t **ipp)
 955 {
 956         if (*ipp == NULL)
 957                 return (0);
 958         /* XXX */
 959         return (0);
 960 }
 961
 962 /*
 963  * When presented with a (*chainp) representing an inode of type
 964  * OBJTYPE_HARDLINK this code will save the original inode (with a ref)
 965  * in (*ipp), and then locate the hidden hardlink target in (dip) or
 966  * any parent directory above (dip).  The locked (*chainp) is replaced
 967  * with a new locked (*chainp) representing the hardlink target.
 968  */
 969 int
 970 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
 971                       hammer2_inode_t **ipp)
 972 {
 973         hammer2_mount_t *hmp = dip->hmp;
 974         hammer2_chain_t *chain = *chainp;
 975         hammer2_chain_t *parent;
 976         hammer2_inode_t *pip;
 977         hammer2_key_t lhc;
 978
 979         *ipp = chain->u.ip;
 980         hammer2_inode_ref(chain->u.ip);
 981         lhc = chain->u.ip->ip_data.inum;
 982
 983         hammer2_inode_unlock_ex(chain->u.ip);
 984         pip = chain->u.ip->pip;
 985
 986         chain = NULL;
 987         while (pip) {
 988                 parent = &pip->chain;
 989                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
 990
 991                 hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
 992                 chain = hammer2_chain_lookup(hmp, &parent, lhc, lhc, 0);
 993                 hammer2_chain_unlock(hmp, parent);
 994                 if (chain)
 995                         break;
 996                 pip = pip->pip; /* XXX SMP RACE */
 997         }
 998         *chainp = chain;
 999         if (chain) {
1000                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1001                 /* already locked */
1002                 return (0);
1003         } else {
1004                 return (EIO);
1005         }
1006 }
1007
1008 /*
1009  * Find the directory common to both fdip and tdip, hold and return
1010  * its inode.
1011  */
1012 hammer2_inode_t *
1013 hammer2_inode_common_parent(hammer2_mount_t *hmp,
1014                             hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1015 {
1016         hammer2_inode_t *scan1;
1017         hammer2_inode_t *scan2;
1018
1019         /*
1020          * We used to have a depth field but it complicated matters too
1021          * much for directory renames.  So now its ugly.  Check for
1022          * simple cases before giving up and doing it the expensive way.
1023          *
1024          * XXX need a bottom-up topology stability lock
1025          */
1026         if (fdip == tdip || fdip == tdip->pip) {
1027                 hammer2_inode_ref(fdip);
1028                 return(fdip);
1029         }
1030         if (fdip->pip == tdip) {
1031                 hammer2_inode_ref(tdip);
1032                 return(tdip);
1033         }
1034         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1035                 scan2 = tdip;
1036                 while (scan2->pmp == tdip->pmp) {
1037                         if (scan1 == scan2) {
1038                                 hammer2_inode_ref(scan1);
1039                                 return(scan1);
1040                         }
1041                         scan2 = scan2->pip;
1042                 }
1043         }
1044         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1045               fdip, tdip);
1046         /* NOT REACHED */
1047         return(NULL);
1048 }