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 /*
  58  * HAMMER2 inode locks
  59  *
  60  * HAMMER2 offers shared locks and exclusive locks on inodes.
  61  *
  62  * An inode's ip->chain pointer is resolved and stable while an inode is
  63  * locked, and can be cleaned out at any time (become NULL) when an inode
  64  * is not locked.
  65  *
  66  * This function handles duplication races and hardlink replacement races
  67  * which can cause ip's cached chain to become stale.
  68  *
  69  * The underlying chain is also locked and returned.
  70  *
  71  * NOTE: We don't combine the inode/chain lock because putting away an
  72  *       inode would otherwise confuse multiple lock holders of the inode.
  73  */
  74 hammer2_chain_t *
  75 hammer2_inode_lock_ex(hammer2_inode_t *ip)
  76 {
  77         hammer2_chain_t *chain;
  78         hammer2_chain_t *ochain;
  79         hammer2_chain_core_t *core;
  80         int error;
  81
  82         hammer2_inode_ref(ip);
  83         ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
  84
  85         chain = ip->chain;
  86         core = chain->core;
  87         for (;;) {
  88                 if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
  89                         spin_lock(&core->cst.spin);
  90                         while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
  91                                 chain = TAILQ_NEXT(chain, core_entry);
  92                         hammer2_chain_ref(chain);
  93                         spin_unlock(&core->cst.spin);
  94                         hammer2_inode_repoint(ip, NULL, chain);
  95                         hammer2_chain_drop(chain);
  96                 }
  97                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
  98                 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
  99                         break;
 100                 hammer2_chain_unlock(chain);
 101         }
 102         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
 103                 error = hammer2_hardlink_find(ip->pip, &chain, &ochain);
 104                 hammer2_chain_drop(ochain);
 105                 KKASSERT(error == 0);
 106                 /* XXX error handling */
 107         }
 108         return (chain);
 109 }
 110
 111 void
 112 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
 113 {
 114         /*
 115          * XXX this will catch parent directories too which we don't
 116          *     really want.
 117          */
 118         if (chain)
 119                 hammer2_chain_unlock(chain);
 120         ccms_thread_unlock(&ip->topo_cst);
 121         hammer2_inode_drop(ip);
 122 }
 123
 124 /*
 125  * NOTE: We don't combine the inode/chain lock because putting away an
 126  *       inode would otherwise confuse multiple lock holders of the inode.
 127  *
 128  *       Shared locks are especially sensitive to having too many shared
 129  *       lock counts (from the same thread) on certain paths which might
 130  *       need to upgrade them.  Only one count of a shared lock can be
 131  *       upgraded.
 132  */
 133 hammer2_chain_t *
 134 hammer2_inode_lock_sh(hammer2_inode_t *ip)
 135 {
 136         hammer2_chain_t *chain;
 137
 138         hammer2_inode_ref(ip);
 139         for (;;) {
 140                 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
 141
 142                 chain = ip->chain;
 143                 KKASSERT(chain != NULL);        /* for now */
 144                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
 145                                           HAMMER2_RESOLVE_SHARED);
 146
 147                 /*
 148                  * Resolve duplication races, resolve hardlinks by giving
 149                  * up and cycling an exclusive lock.
 150                  */
 151                 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0 &&
 152                     chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK) {
 153                         break;
 154                 }
 155                 hammer2_chain_unlock(chain);
 156                 ccms_thread_unlock(&ip->topo_cst);
 157                 chain = hammer2_inode_lock_ex(ip);
 158                 hammer2_inode_unlock_ex(ip, chain);
 159         }
 160         return (chain);
 161 }
 162
 163 void
 164 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
 165 {
 166         if (chain)
 167                 hammer2_chain_unlock(chain);
 168         ccms_thread_unlock(&ip->topo_cst);
 169         hammer2_inode_drop(ip);
 170 }
 171
 172 ccms_state_t
 173 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
 174 {
 175         return(ccms_thread_lock_temp_release(&ip->topo_cst));
 176 }
 177
 178 void
 179 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
 180 {
 181         ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
 182 }
 183
 184 ccms_state_t
 185 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
 186 {
 187         return(ccms_thread_lock_upgrade(&ip->topo_cst));
 188 }
 189
 190 void
 191 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
 192 {
 193         ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
 194 }
 195
 196 /*
 197  * Lookup an inode by inode number
 198  */
 199 hammer2_inode_t *
 200 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
 201 {
 202         hammer2_inode_t *ip;
 203
 204         if (pmp) {
 205                 spin_lock(&pmp->inum_spin);
 206                 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
 207                 if (ip)
 208                         hammer2_inode_ref(ip);
 209                 spin_unlock(&pmp->inum_spin);
 210         } else {
 211                 ip = NULL;
 212         }
 213         return(ip);
 214 }
 215
 216 /*
 217  * Adding a ref to an inode is only legal if the inode already has at least
 218  * one ref.
 219  */
 220 void
 221 hammer2_inode_ref(hammer2_inode_t *ip)
 222 {
 223         atomic_add_int(&ip->refs, 1);
 224 }
 225
 226 /*
 227  * Drop an inode reference, freeing the inode when the last reference goes
 228  * away.
 229  */
 230 void
 231 hammer2_inode_drop(hammer2_inode_t *ip)
 232 {
 233         hammer2_pfsmount_t *pmp;
 234         hammer2_inode_t *pip;
 235         u_int refs;
 236
 237         while (ip) {
 238                 refs = ip->refs;
 239                 cpu_ccfence();
 240                 if (refs == 1) {
 241                         /*
 242                          * Transition to zero, must interlock with
 243                          * the inode inumber lookup tree (if applicable).
 244                          *
 245                          * NOTE: The super-root inode has no pmp.
 246                          */
 247                         pmp = ip->pmp;
 248                         if (pmp)
 249                                 spin_lock(&pmp->inum_spin);
 250
 251                         if (atomic_cmpset_int(&ip->refs, 1, 0)) {
 252                                 KKASSERT(ip->topo_cst.count == 0);
 253                                 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
 254                                         atomic_clear_int(&ip->flags,
 255                                                      HAMMER2_INODE_ONRBTREE);
 256                                         RB_REMOVE(hammer2_inode_tree,
 257                                                   &pmp->inum_tree, ip);
 258                                 }
 259                                 if (pmp)
 260                                         spin_unlock(&pmp->inum_spin);
 261
 262                                 pip = ip->pip;
 263                                 ip->pip = NULL;
 264                                 ip->pmp = NULL;
 265
 266                                 /*
 267                                  * Cleaning out ip->chain isn't entirely
 268                                  * trivial.
 269                                  */
 270                                 hammer2_inode_repoint(ip, NULL, NULL);
 271
 272                                 /*
 273                                  * We have to drop pip (if non-NULL) to
 274                                  * dispose of our implied reference from
 275                                  * ip->pip.  We can simply loop on it.
 276                                  */
 277                                 if (pmp) {
 278                                         KKASSERT((ip->flags &
 279                                                   HAMMER2_INODE_SROOT) == 0);
 280                                         kfree(ip, pmp->minode);
 281                                         atomic_add_long(&pmp->inmem_inodes, -1);
 282                                 } else {
 283                                         KKASSERT(ip->flags &
 284                                                  HAMMER2_INODE_SROOT);
 285                                         kfree(ip, M_HAMMER2);
 286                                 }
 287                                 ip = pip;
 288                                 /* continue with pip (can be NULL) */
 289                         } else {
 290                                 if (pmp)
 291                                         spin_unlock(&ip->pmp->inum_spin);
 292                         }
 293                 } else {
 294                         /*
 295                          * Non zero transition
 296                          */
 297                         if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
 298                                 break;
 299                 }
 300         }
 301 }
 302
 303 /*
 304  * Get the vnode associated with the given inode, allocating the vnode if
 305  * necessary.  The vnode will be returned exclusively locked.
 306  *
 307  * The caller must lock the inode (shared or exclusive).
 308  *
 309  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
 310  * races.
 311  */
 312 struct vnode *
 313 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
 314 {
 315         hammer2_inode_data_t *ipdata;
 316         hammer2_pfsmount_t *pmp;
 317         struct vnode *vp;
 318         ccms_state_t ostate;
 319
 320         pmp = ip->pmp;
 321         KKASSERT(pmp != NULL);
 322         *errorp = 0;
 323         ipdata = &ip->chain->data->ipdata;
 324
 325         for (;;) {
 326                 /*
 327                  * Attempt to reuse an existing vnode assignment.  It is
 328                  * possible to race a reclaim so the vget() may fail.  The
 329                  * inode must be unlocked during the vget() to avoid a
 330                  * deadlock against a reclaim.
 331                  */
 332                 vp = ip->vp;
 333                 if (vp) {
 334                         /*
 335                          * Inode must be unlocked during the vget() to avoid
 336                          * possible deadlocks, but leave the ip ref intact.
 337                          *
 338                          * vnode is held to prevent destruction during the
 339                          * vget().  The vget() can still fail if we lost
 340                          * a reclaim race on the vnode.
 341                          */
 342                         vhold(vp);
 343                         ostate = hammer2_inode_lock_temp_release(ip);
 344                         if (vget(vp, LK_EXCLUSIVE)) {
 345                                 vdrop(vp);
 346                                 hammer2_inode_lock_temp_restore(ip, ostate);
 347                                 continue;
 348                         }
 349                         hammer2_inode_lock_temp_restore(ip, ostate);
 350                         vdrop(vp);
 351                         /* vp still locked and ref from vget */
 352                         if (ip->vp != vp) {
 353                                 kprintf("hammer2: igetv race %p/%p\n",
 354                                         ip->vp, vp);
 355                                 vput(vp);
 356                                 continue;
 357                         }
 358                         *errorp = 0;
 359                         break;
 360                 }
 361
 362                 /*
 363                  * No vnode exists, allocate a new vnode.  Beware of
 364                  * allocation races.  This function will return an
 365                  * exclusively locked and referenced vnode.
 366                  */
 367                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
 368                 if (*errorp) {
 369                         kprintf("hammer2: igetv getnewvnode failed %d\n",
 370                                 *errorp);
 371                         vp = NULL;
 372                         break;
 373                 }
 374
 375                 /*
 376                  * Lock the inode and check for an allocation race.
 377                  */
 378                 ostate = hammer2_inode_lock_upgrade(ip);
 379                 if (ip->vp != NULL) {
 380                         vp->v_type = VBAD;
 381                         vx_put(vp);
 382                         hammer2_inode_lock_downgrade(ip, ostate);
 383                         continue;
 384                 }
 385
 386                 switch (ipdata->type) {
 387                 case HAMMER2_OBJTYPE_DIRECTORY:
 388                         vp->v_type = VDIR;
 389                         break;
 390                 case HAMMER2_OBJTYPE_REGFILE:
 391                         vp->v_type = VREG;
 392                         vinitvmio(vp, ipdata->size,
 393                                   HAMMER2_LBUFSIZE,
 394                                   (int)ipdata->size & HAMMER2_LBUFMASK);
 395                         break;
 396                 case HAMMER2_OBJTYPE_SOFTLINK:
 397                         /*
 398                          * XXX for now we are using the generic file_read
 399                          * and file_write code so we need a buffer cache
 400                          * association.
 401                          */
 402                         vp->v_type = VLNK;
 403                         vinitvmio(vp, ipdata->size,
 404                                   HAMMER2_LBUFSIZE,
 405                                   (int)ipdata->size & HAMMER2_LBUFMASK);
 406                         break;
 407                 case HAMMER2_OBJTYPE_CDEV:
 408                         vp->v_type = VCHR;
 409                         /* fall through */
 410                 case HAMMER2_OBJTYPE_BDEV:
 411                         vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
 412                         if (ipdata->type != HAMMER2_OBJTYPE_CDEV)
 413                                 vp->v_type = VBLK;
 414                         addaliasu(vp, ipdata->rmajor, ipdata->rminor);
 415                         break;
 416                 case HAMMER2_OBJTYPE_FIFO:
 417                         vp->v_type = VFIFO;
 418                         vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
 419                         break;
 420                 default:
 421                         panic("hammer2: unhandled objtype %d", ipdata->type);
 422                         break;
 423                 }
 424
 425                 if (ip == pmp->iroot)
 426                         vsetflags(vp, VROOT);
 427
 428                 vp->v_data = ip;
 429                 ip->vp = vp;
 430                 hammer2_inode_ref(ip);          /* vp association */
 431                 hammer2_inode_lock_downgrade(ip, ostate);
 432                 break;
 433         }
 434
 435         /*
 436          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
 437          */
 438         if (hammer2_debug & 0x0002) {
 439                 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
 440                         vp, vp->v_refcnt, vp->v_auxrefs);
 441         }
 442         return (vp);
 443 }
 444
 445 /*
 446  * The passed-in chain must be locked and the returned inode will also be
 447  * locked.  This routine typically locates or allocates the inode, assigns
 448  * ip->chain (adding a ref to chain if necessary), and returns the inode.
 449  *
 450  * The hammer2_inode structure regulates the interface between the high level
 451  * kernel VNOPS API and the filesystem backend (the chains).
 452  *
 453  * WARNING!  This routine sucks up the chain's lock (makes it part of the
 454  *           inode lock from the point of view of the inode lock API),
 455  *           so callers need to be careful.
 456  *
 457  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
 458  *           is allowed to pass pmp == NULL and dip == NULL for sroot.
 459  */
 460 hammer2_inode_t *
 461 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
 462                   hammer2_chain_t *chain)
 463 {
 464         hammer2_inode_t *nip;
 465
 466         KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
 467
 468         /*
 469          * Interlocked lookup/ref of the inode.  This code is only needed
 470          * when looking up inodes with nlinks != 0 (TODO: optimize out
 471          * otherwise and test for duplicates).
 472          */
 473 again:
 474         for (;;) {
 475                 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
 476                 if (nip == NULL)
 477                         break;
 478                 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
 479                 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
 480                         ccms_thread_unlock(&nip->topo_cst);
 481                         hammer2_inode_drop(nip);
 482                         continue;
 483                 }
 484                 if (nip->chain != chain)
 485                         hammer2_inode_repoint(nip, NULL, chain);
 486
 487                 /*
 488                  * Consolidated nip/nip->chain is locked (chain locked
 489                  * by caller).
 490                  */
 491                 return nip;
 492         }
 493
 494         /*
 495          * We couldn't find the inode number, create a new inode.
 496          */
 497         if (pmp) {
 498                 nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
 499                 atomic_add_long(&pmp->inmem_inodes, 1);
 500                 hammer2_chain_memory_inc(pmp);
 501                 hammer2_chain_memory_wakeup(pmp);
 502         } else {
 503                 nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
 504                 nip->flags = HAMMER2_INODE_SROOT;
 505         }
 506         nip->inum = chain->data->ipdata.inum;
 507         nip->size = chain->data->ipdata.size;
 508         nip->mtime = chain->data->ipdata.mtime;
 509         hammer2_inode_repoint(nip, NULL, chain);
 510         nip->pip = dip;                         /* can be NULL */
 511         if (dip)
 512                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
 513
 514         nip->pmp = pmp;
 515
 516         /*
 517          * ref and lock on nip gives it state compatible to after a
 518          * hammer2_inode_lock_ex() call.
 519          */
 520         nip->refs = 1;
 521         ccms_cst_init(&nip->topo_cst, &nip->chain);
 522         ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
 523         /* combination of thread lock and chain lock == inode lock */
 524
 525         /*
 526          * Attempt to add the inode.  If it fails we raced another inode
 527          * get.  Undo all the work and try again.
 528          */
 529         if (pmp) {
 530                 spin_lock(&pmp->inum_spin);
 531                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
 532                         spin_unlock(&pmp->inum_spin);
 533                         ccms_thread_unlock(&nip->topo_cst);
 534                         hammer2_inode_drop(nip);
 535                         goto again;
 536                 }
 537                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
 538                 spin_unlock(&pmp->inum_spin);
 539         }
 540
 541         return (nip);
 542 }
 543
 544 /*
 545  * Create a new inode in the specified directory using the vattr to
 546  * figure out the type of inode.
 547  *
 548  * If no error occurs the new inode with its chain locked is returned in
 549  * *nipp, otherwise an error is returned and *nipp is set to NULL.
 550  *
 551  * If vap and/or cred are NULL the related fields are not set and the
 552  * inode type defaults to a directory.  This is used when creating PFSs
 553  * under the super-root, so the inode number is set to 1 in this case.
 554  *
 555  * dip is not locked on entry.
 556  */
 557 hammer2_inode_t *
 558 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
 559                      struct vattr *vap, struct ucred *cred,
 560                      const uint8_t *name, size_t name_len,
 561                      hammer2_chain_t **chainp, int *errorp)
 562 {
 563         hammer2_inode_data_t *dipdata;
 564         hammer2_inode_data_t *nipdata;
 565         hammer2_chain_t *chain;
 566         hammer2_chain_t *parent;
 567         hammer2_inode_t *nip;
 568         hammer2_key_t key_dummy;
 569         hammer2_key_t lhc;
 570         int error;
 571         uid_t xuid;
 572         uuid_t dip_uid;
 573         uuid_t dip_gid;
 574         uint32_t dip_mode;
 575         uint8_t dip_algo;
 576         int cache_index = -1;
 577
 578         lhc = hammer2_dirhash(name, name_len);
 579         *errorp = 0;
 580
 581         /*
 582          * Locate the inode or indirect block to create the new
 583          * entry in.  At the same time check for key collisions
 584          * and iterate until we don't get one.
 585          *
 586          * NOTE: hidden inodes do not have iterators.
 587          */
 588 retry:
 589         parent = hammer2_inode_lock_ex(dip);
 590         dipdata = &dip->chain->data->ipdata;
 591         dip_uid = dipdata->uid;
 592         dip_gid = dipdata->gid;
 593         dip_mode = dipdata->mode;
 594         dip_algo = dipdata->comp_algo;
 595
 596         error = 0;
 597         while (error == 0) {
 598                 chain = hammer2_chain_lookup(&parent, &key_dummy,
 599                                              lhc, lhc, &cache_index, 0);
 600                 if (chain == NULL)
 601                         break;
 602                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
 603                         error = ENOSPC;
 604                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 605                         error = ENOSPC;
 606                 hammer2_chain_unlock(chain);
 607                 chain = NULL;
 608                 ++lhc;
 609         }
 610
 611         if (error == 0) {
 612                 error = hammer2_chain_create(trans, &parent, &chain,
 613                                              lhc, 0,
 614                                              HAMMER2_BREF_TYPE_INODE,
 615                                              HAMMER2_INODE_BYTES);
 616         }
 617
 618         /*
 619          * Cleanup and handle retries.
 620          */
 621         if (error == EAGAIN) {
 622                 hammer2_chain_ref(parent);
 623                 hammer2_inode_unlock_ex(dip, parent);
 624                 hammer2_chain_wait(parent);
 625                 hammer2_chain_drop(parent);
 626                 goto retry;
 627         }
 628         hammer2_inode_unlock_ex(dip, parent);
 629
 630         if (error) {
 631                 KKASSERT(chain == NULL);
 632                 *errorp = error;
 633                 return (NULL);
 634         }
 635
 636         /*
 637          * Set up the new inode.
 638          *
 639          * NOTE: *_get() integrates chain's lock into the inode lock.
 640          *
 641          * NOTE: Only one new inode can currently be created per
 642          *       transaction.  If the need arises we can adjust
 643          *       hammer2_trans_init() to allow more.
 644          *
 645          * NOTE: nipdata will have chain's blockset data.
 646          */
 647         chain->data->ipdata.inum = trans->inode_tid;
 648         nip = hammer2_inode_get(dip->pmp, dip, chain);
 649         nipdata = &chain->data->ipdata;
 650
 651         if (vap) {
 652                 KKASSERT(trans->inodes_created == 0);
 653                 nipdata->type = hammer2_get_obj_type(vap->va_type);
 654                 nipdata->inum = trans->inode_tid;
 655                 ++trans->inodes_created;
 656
 657                 switch (nipdata->type) {
 658                 case HAMMER2_OBJTYPE_CDEV:
 659                 case HAMMER2_OBJTYPE_BDEV:
 660                         nipdata->rmajor = vap->va_rmajor;
 661                         nipdata->rminor = vap->va_rminor;
 662                         break;
 663                 default:
 664                         break;
 665                 }
 666         } else {
 667                 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
 668                 nipdata->inum = 1;
 669         }
 670
 671         /* Inherit parent's inode compression mode. */
 672         nip->comp_heuristic = 0;
 673         nipdata->comp_algo = dip_algo;
 674         nipdata->version = HAMMER2_INODE_VERSION_ONE;
 675         hammer2_update_time(&nipdata->ctime);
 676         nipdata->mtime = nipdata->ctime;
 677         if (vap)
 678                 nipdata->mode = vap->va_mode;
 679         nipdata->nlinks = 1;
 680         if (vap) {
 681                 if (dip && dip->pmp) {
 682                         xuid = hammer2_to_unix_xid(&dip_uid);
 683                         xuid = vop_helper_create_uid(dip->pmp->mp,
 684                                                      dip_mode,
 685                                                      xuid,
 686                                                      cred,
 687                                                      &vap->va_mode);
 688                 } else {
 689                         /* super-root has no dip and/or pmp */
 690                         xuid = 0;
 691                 }
 692                 if (vap->va_vaflags & VA_UID_UUID_VALID)
 693                         nipdata->uid = vap->va_uid_uuid;
 694                 else if (vap->va_uid != (uid_t)VNOVAL)
 695                         hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
 696                 else
 697                         hammer2_guid_to_uuid(&nipdata->uid, xuid);
 698
 699                 if (vap->va_vaflags & VA_GID_UUID_VALID)
 700                         nipdata->gid = vap->va_gid_uuid;
 701                 else if (vap->va_gid != (gid_t)VNOVAL)
 702                         hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
 703                 else if (dip)
 704                         nipdata->gid = dip_gid;
 705         }
 706
 707         /*
 708          * Regular files and softlinks allow a small amount of data to be
 709          * directly embedded in the inode.  This flag will be cleared if
 710          * the size is extended past the embedded limit.
 711          */
 712         if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
 713             nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
 714                 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
 715         }
 716
 717         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
 718         bcopy(name, nipdata->filename, name_len);
 719         nipdata->name_key = lhc;
 720         nipdata->name_len = name_len;
 721         *chainp = chain;
 722
 723         return (nip);
 724 }
 725
 726 /*
 727  * chain may have been moved around by the create.
 728  */
 729 void
 730 hammer2_chain_refactor(hammer2_chain_t **chainp)
 731 {
 732         hammer2_chain_t *chain = *chainp;
 733         hammer2_chain_core_t *core;
 734
 735         core = chain->core;
 736         while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
 737                 spin_lock(&core->cst.spin);
 738                 chain = TAILQ_NEXT(chain, core_entry);
 739                 while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
 740                         chain = TAILQ_NEXT(chain, core_entry);
 741                 hammer2_chain_ref(chain);
 742                 spin_unlock(&core->cst.spin);
 743                 KKASSERT(chain->core == core);
 744
 745                 hammer2_chain_unlock(*chainp);
 746                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
 747                                           HAMMER2_RESOLVE_NOREF); /* eat ref */
 748                 *chainp = chain;
 749         }
 750 }
 751
 752 /*
 753  * Shift *chainp up to the specified directory, change the filename
 754  * to "0xINODENUMBER", and adjust the key.  The chain becomes the
 755  * invisible hardlink target.
 756  *
 757  * The original *chainp has already been marked deleted.
 758  */
 759 static
 760 void
 761 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 762                         hammer2_inode_t *dip, int nlinks, int *errorp)
 763 {
 764         hammer2_inode_data_t *nipdata;
 765         hammer2_chain_t *parent;
 766         hammer2_chain_t *chain;
 767         hammer2_chain_t *xchain;
 768         hammer2_key_t key_dummy;
 769         hammer2_key_t lhc;
 770         hammer2_blockref_t bref;
 771         int cache_index = -1;
 772
 773         chain = *chainp;
 774         lhc = chain->data->ipdata.inum;
 775         KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
 776
 777         /*
 778          * Locate the inode or indirect block to create the new
 779          * entry in.  lhc represents the inode number so there is
 780          * no collision iteration.
 781          *
 782          * There should be no key collisions with invisible inode keys.
 783          *
 784          * WARNING! Must use inode_lock_ex() on dip to handle a stale
 785          *          dip->chain cache.
 786          */
 787 retry:
 788         *errorp = 0;
 789         parent = hammer2_inode_lock_ex(dip);
 790         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
 791         xchain = hammer2_chain_lookup(&parent, &key_dummy,
 792                                       lhc, lhc, &cache_index, 0);
 793         if (xchain) {
 794                 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
 795                         xchain, parent, dip, dip->chain);
 796                 hammer2_chain_unlock(xchain);
 797                 xchain = NULL;
 798                 *errorp = ENOSPC;
 799 #if 0
 800                 Debugger("X3");
 801 #endif
 802         }
 803
 804         /*
 805          * Create entry in common parent directory using the seek position
 806          * calculated above.
 807          *
 808          * We must refactor chain because it might have been shifted into
 809          * an indirect chain by the create.
 810          */
 811         if (*errorp == 0) {
 812                 KKASSERT(xchain == NULL);
 813                 *errorp = hammer2_chain_create(trans, &parent, &xchain,
 814                                                lhc, 0,
 815                                                HAMMER2_BREF_TYPE_INODE,/* n/a */
 816                                                HAMMER2_INODE_BYTES);   /* n/a */
 817                 /*XXX this somehow isn't working on chain XXX*/
 818                 /*KKASSERT(xxx)*/
 819         }
 820
 821         /*
 822          * Cleanup and handle retries.
 823          */
 824         if (*errorp == EAGAIN) {
 825                 hammer2_chain_ref(parent);
 826                 /* hammer2_chain_lookup_done(parent); */
 827                 hammer2_inode_unlock_ex(dip, parent);
 828                 hammer2_chain_wait(parent);
 829                 hammer2_chain_drop(parent);
 830                 goto retry;
 831         }
 832
 833         /*
 834          * Handle the error case
 835          */
 836         if (*errorp) {
 837                 panic("error2");
 838                 KKASSERT(xchain == NULL);
 839                 hammer2_inode_unlock_ex(dip, parent);
 840                 /*hammer2_chain_lookup_done(parent);*/
 841                 return;
 842         }
 843
 844         /*
 845          * Use xchain as a placeholder for (lhc).  Duplicate chain to the
 846          * same target bref as xchain and then delete xchain.  The duplication
 847          * occurs after xchain in flush order even though xchain is deleted
 848          * after the duplication. XXX
 849          *
 850          * WARNING! Duplications (to a different parent) can cause indirect
 851          *          blocks to be inserted, refactor xchain.
 852          */
 853         bref = chain->bref;
 854         bref.key = lhc;                 /* invisible dir entry key */
 855         bref.keybits = 0;
 856         hammer2_chain_delete(trans, xchain, 0);
 857         hammer2_chain_duplicate(trans, &parent, &chain, &bref, 0, 2);
 858         hammer2_chain_refactor(&xchain);
 859         /*hammer2_chain_delete(trans, xchain, 0);*/
 860
 861         hammer2_inode_unlock_ex(dip, parent);
 862         /*hammer2_chain_lookup_done(parent);*/
 863         hammer2_chain_unlock(xchain);   /* no longer needed */
 864
 865         /*
 866          * chain is now 'live' again.. adjust the filename.
 867          *
 868          * Directory entries are inodes but this is a hidden hardlink
 869          * target.  The name isn't used but to ease debugging give it
 870          * a name after its inode number.
 871          */
 872         hammer2_chain_modify(trans, &chain, 0);
 873         nipdata = &chain->data->ipdata;
 874         ksnprintf(nipdata->filename, sizeof(nipdata->filename),
 875                   "0x%016jx", (intmax_t)nipdata->inum);
 876         nipdata->name_len = strlen(nipdata->filename);
 877         nipdata->name_key = lhc;
 878         nipdata->nlinks += nlinks;
 879
 880         *chainp = chain;
 881 }
 882
 883 /*
 884  * Connect the target inode represented by (*chainp) to the media topology
 885  * at (dip, name, len).
 886  *
 887  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
 888  * entry instead of connecting (*chainp).
 889  *
 890  * If hlink is FALSE this function uses chain_duplicate() to make a copy
 891  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
 892  * by the caller in this case (e.g. rename).
 893  */
 894 int
 895 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
 896                       hammer2_inode_t *dip, hammer2_chain_t **chainp,
 897                       const uint8_t *name, size_t name_len)
 898 {
 899         hammer2_inode_data_t *ipdata;
 900         hammer2_chain_t *nchain;
 901         hammer2_chain_t *parent;
 902         hammer2_chain_t *ochain;
 903         hammer2_key_t key_dummy;
 904         hammer2_key_t lhc;
 905         int cache_index = -1;
 906         int error;
 907
 908         ochain = *chainp;
 909
 910         /*
 911          * Since ochain is either disconnected from the topology or represents
 912          * a hardlink terminus which is always a parent of or equal to dip,
 913          * we should be able to safely lock dip->chain for our setup.
 914          *
 915          * WARNING! Must use inode_lock_ex() on dip to handle a stale
 916          *          dip->chain cache.
 917          */
 918         parent = hammer2_inode_lock_ex(dip);
 919         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
 920
 921         lhc = hammer2_dirhash(name, name_len);
 922
 923         /*
 924          * Locate the inode or indirect block to create the new
 925          * entry in.  At the same time check for key collisions
 926          * and iterate until we don't get one.
 927          */
 928         error = 0;
 929         while (error == 0) {
 930                 nchain = hammer2_chain_lookup(&parent, &key_dummy,
 931                                               lhc, lhc, &cache_index, 0);
 932                 if (nchain == NULL)
 933                         break;
 934                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
 935                         error = ENOSPC;
 936                 hammer2_chain_unlock(nchain);
 937                 nchain = NULL;
 938                 ++lhc;
 939         }
 940
 941         if (error == 0) {
 942                 if (hlink) {
 943                         /*
 944                          * Hardlink pointer needed, create totally fresh
 945                          * directory entry.
 946                          *
 947                          * We must refactor ochain because it might have
 948                          * been shifted into an indirect chain by the
 949                          * create.
 950                          */
 951                         KKASSERT(nchain == NULL);
 952                         error = hammer2_chain_create(trans, &parent, &nchain,
 953                                                      lhc, 0,
 954                                                      HAMMER2_BREF_TYPE_INODE,
 955                                                      HAMMER2_INODE_BYTES);
 956                         hammer2_chain_refactor(&ochain);
 957                 } else {
 958                         /*
 959                          * Reconnect the original chain and rename.  Use
 960                          * chain_duplicate().  The caller will likely delete
 961                          * or has already deleted the original chain in
 962                          * this case.
 963                          *
 964                          * NOTE: chain_duplicate() generates a new chain
 965                          *       with CHAIN_DELETED cleared (ochain typically
 966                          *       has it set from the file unlink).
 967                          *
 968                          * WARNING! Can cause held-over chains to require a
 969                          *          refactor.  Fortunately we have none (our
 970                          *          locked chains are passed into and
 971                          *          modified by the call).
 972                          */
 973                         nchain = ochain;
 974                         ochain = NULL;
 975                         hammer2_chain_duplicate(trans, NULL, &nchain, NULL,
 976                                                 0, 3);
 977                         error = hammer2_chain_create(trans, &parent, &nchain,
 978                                                      lhc, 0,
 979                                                      HAMMER2_BREF_TYPE_INODE,
 980                                                      HAMMER2_INODE_BYTES);
 981                 }
 982         }
 983
 984         /*
 985          * Unlock stuff.
 986          */
 987         KKASSERT(error != EAGAIN);
 988         hammer2_inode_unlock_ex(dip, parent);
 989         /*hammer2_chain_lookup_done(parent);*/
 990         parent = NULL;
 991
 992         /*
 993          * nchain should be NULL on error, leave ochain (== *chainp) alone.
 994          */
 995         if (error) {
 996                 KKASSERT(nchain == NULL);
 997                 return (error);
 998         }
 999
1000         /*
1001          * Directory entries are inodes so if the name has changed we have
1002          * to update the inode.
1003          *
1004          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
1005          * chain, the caller will access the hardlink via the actual hardlink
1006          * target file and not the hardlink pointer entry, so we must still
1007          * return ochain.
1008          */
1009         if (hlink && hammer2_hardlink_enable >= 0) {
1010                 /*
1011                  * Create the HARDLINK pointer.  oip represents the hardlink
1012                  * target in this situation.
1013                  *
1014                  * We will return ochain (the hardlink target).
1015                  */
1016                 hammer2_chain_modify(trans, &nchain, 0);
1017                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1018                 ipdata = &nchain->data->ipdata;
1019                 bcopy(name, ipdata->filename, name_len);
1020                 ipdata->name_key = lhc;
1021                 ipdata->name_len = name_len;
1022                 ipdata->target_type = ochain->data->ipdata.type;
1023                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1024                 ipdata->inum = ochain->data->ipdata.inum;
1025                 ipdata->nlinks = 1;
1026                 hammer2_chain_unlock(nchain);
1027                 nchain = ochain;
1028                 ochain = NULL;
1029         } else if (hlink && hammer2_hardlink_enable < 0) {
1030                 /*
1031                  * Create a snapshot (hardlink fake mode for debugging).
1032                  * (ochain already flushed above so we can just copy the
1033                  * bref XXX).
1034                  *
1035                  * Since this is a snapshot we return nchain in the fake
1036                  * hardlink case.
1037                  */
1038                 hammer2_chain_modify(trans, &nchain, 0);
1039                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1040                 ipdata = &nchain->data->ipdata;
1041                 *ipdata = ochain->data->ipdata;
1042                 bcopy(name, ipdata->filename, name_len);
1043                 ipdata->name_key = lhc;
1044                 ipdata->name_len = name_len;
1045                 atomic_clear_int(&nchain->core->flags,
1046                                  HAMMER2_CORE_COUNTEDBREFS);
1047                 kprintf("created fake hardlink %*.*s\n",
1048                         (int)name_len, (int)name_len, name);
1049         } else {
1050                 /*
1051                  * nchain is a duplicate of ochain at the new location.
1052                  * We must fixup the name stored in oip.  The bref key
1053                  * has already been set up.
1054                  */
1055                 hammer2_chain_modify(trans, &nchain, 0);
1056                 ipdata = &nchain->data->ipdata;
1057
1058                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1059                 bcopy(name, ipdata->filename, name_len);
1060                 ipdata->name_key = lhc;
1061                 ipdata->name_len = name_len;
1062                 ipdata->nlinks = 1;
1063         }
1064
1065         /*
1066          * We are replacing ochain with nchain, unlock ochain.  In the
1067          * case where ochain is left unchanged the code above sets
1068          * nchain to ochain and ochain to NULL, resulting in a NOP here.
1069          */
1070         if (ochain)
1071                 hammer2_chain_unlock(ochain);
1072         *chainp = nchain;
1073
1074         return (0);
1075 }
1076
1077 /*
1078  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1079  * locked.
1080  *
1081  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1082  * and nchain is ref'd.
1083  */
1084 void
1085 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1086                       hammer2_chain_t *nchain)
1087 {
1088         hammer2_chain_t *ochain;
1089         hammer2_inode_t *opip;
1090
1091         /*
1092          * Repoint ip->chain if requested.
1093          */
1094         ochain = ip->chain;
1095         ip->chain = nchain;
1096         if (nchain)
1097                 hammer2_chain_ref(nchain);
1098         if (ochain)
1099                 hammer2_chain_drop(ochain);
1100
1101         /*
1102          * Repoint ip->pip if requested (non-NULL pip).
1103          */
1104         if (pip && ip->pip != pip) {
1105                 opip = ip->pip;
1106                 hammer2_inode_ref(pip);
1107                 ip->pip = pip;
1108                 if (opip)
1109                         hammer2_inode_drop(opip);
1110         }
1111 }
1112
1113 /*
1114  * Unlink the file from the specified directory inode.  The directory inode
1115  * does not need to be locked.
1116  *
1117  * isdir determines whether a directory/non-directory check should be made.
1118  * No check is made if isdir is set to -1.
1119  *
1120  * NOTE!  The underlying file can still be active with open descriptors
1121  *        or if the chain is being manually held (e.g. for rename).
1122  *
1123  *        The caller is responsible for fixing up ip->chain if e.g. a
1124  *        rename occurs (see chain_duplicate()).
1125  */
1126 int
1127 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1128                     const uint8_t *name, size_t name_len,
1129                     int isdir, int *hlinkp)
1130 {
1131         hammer2_inode_data_t *ipdata;
1132         hammer2_chain_t *parent;
1133         hammer2_chain_t *ochain;
1134         hammer2_chain_t *chain;
1135         hammer2_chain_t *dparent;
1136         hammer2_chain_t *dchain;
1137         hammer2_key_t key_dummy;
1138         hammer2_key_t key_next;
1139         hammer2_key_t lhc;
1140         int error;
1141         int cache_index = -1;
1142         uint8_t type;
1143
1144         error = 0;
1145         ochain = NULL;
1146         lhc = hammer2_dirhash(name, name_len);
1147
1148         /*
1149          * Search for the filename in the directory
1150          */
1151         if (hlinkp)
1152                 *hlinkp = 0;
1153         parent = hammer2_inode_lock_ex(dip);
1154         chain = hammer2_chain_lookup(&parent, &key_next,
1155                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1156                                      &cache_index, 0);
1157         while (chain) {
1158                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1159                     name_len == chain->data->ipdata.name_len &&
1160                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1161                         break;
1162                 }
1163                 chain = hammer2_chain_next(&parent, chain, &key_next,
1164                                            key_next,
1165                                            lhc + HAMMER2_DIRHASH_LOMASK,
1166                                            &cache_index, 0);
1167         }
1168         hammer2_inode_unlock_ex(dip, NULL);     /* retain parent */
1169
1170         /*
1171          * Not found or wrong type (isdir < 0 disables the type check).
1172          * If a hardlink pointer, type checks use the hardlink target.
1173          */
1174         if (chain == NULL) {
1175                 error = ENOENT;
1176                 goto done;
1177         }
1178         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1179                 if (hlinkp)
1180                         *hlinkp = 1;
1181                 type = chain->data->ipdata.target_type;
1182         }
1183
1184         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1185                 error = ENOTDIR;
1186                 goto done;
1187         }
1188         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1189                 error = EISDIR;
1190                 goto done;
1191         }
1192
1193         /*
1194          * Hardlink must be resolved.  We can't hold parent locked while we
1195          * do this or we could deadlock.
1196          *
1197          * On success chain will be adjusted to point at the hardlink target
1198          * and ochain will point to the hardlink pointer in the original
1199          * directory.  Otherwise chain remains pointing to the original.
1200          */
1201         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1202                 hammer2_chain_unlock(parent);
1203                 parent = NULL;
1204                 error = hammer2_hardlink_find(dip, &chain, &ochain);
1205         }
1206
1207         /*
1208          * If this is a directory the directory must be empty.  However, if
1209          * isdir < 0 we are doing a rename and the directory does not have
1210          * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1211          * and the directory does not have to be empty.
1212          *
1213          * NOTE: We check the full key range here which covers both visible
1214          *       and invisible entries.  Theoretically there should be no
1215          *       invisible (hardlink target) entries if there are no visible
1216          *       entries.
1217          */
1218         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1219                 dparent = hammer2_chain_lookup_init(chain, 0);
1220                 dchain = hammer2_chain_lookup(&dparent, &key_dummy,
1221                                               0, (hammer2_key_t)-1,
1222                                               &cache_index,
1223                                               HAMMER2_LOOKUP_NODATA);
1224                 if (dchain) {
1225                         hammer2_chain_unlock(dchain);
1226                         hammer2_chain_lookup_done(dparent);
1227                         error = ENOTEMPTY;
1228                         goto done;
1229                 }
1230                 hammer2_chain_lookup_done(dparent);
1231                 dparent = NULL;
1232                 /* dchain NULL */
1233         }
1234
1235         /*
1236          * Ok, we can now unlink the chain.  We always decrement nlinks even
1237          * if the entry can be deleted in case someone has the file open and
1238          * does an fstat().
1239          *
1240          * The chain itself will no longer be in the on-media topology but
1241          * can still be flushed to the media (e.g. if an open descriptor
1242          * remains).  When the last vnode/ip ref goes away the chain will
1243          * be marked unmodified, avoiding any further (now unnecesary) I/O.
1244          *
1245          * A non-NULL ochain indicates a hardlink.
1246          */
1247         if (ochain) {
1248                 /*
1249                  * Delete the original hardlink pointer.
1250                  *
1251                  * NOTE: parent from above is NULL when ochain != NULL
1252                  *       so we can reuse it.
1253                  */
1254                 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1255                 hammer2_chain_delete(trans, ochain, 0);
1256                 hammer2_chain_unlock(ochain);
1257
1258                 /*
1259                  * Then decrement nlinks on hardlink target, deleting
1260                  * the target when nlinks drops to 0.
1261                  */
1262                 hammer2_chain_modify(trans, &chain, 0);
1263                 --chain->data->ipdata.nlinks;
1264                 if (chain->data->ipdata.nlinks == 0)
1265                         hammer2_chain_delete(trans, chain, 0);
1266         } else {
1267                 /*
1268                  * Otherwise this was not a hardlink and we can just
1269                  * remove the entry and decrement nlinks.
1270                  *
1271                  * NOTE: *_get() integrates chain's lock into the inode lock.
1272                  */
1273                 hammer2_chain_modify(trans, &chain, 0);
1274                 ipdata = &chain->data->ipdata;
1275                 --ipdata->nlinks;
1276                 hammer2_chain_delete(trans, chain, 0);
1277         }
1278
1279         error = 0;
1280 done:
1281         if (chain)
1282                 hammer2_chain_unlock(chain);
1283         if (parent)
1284                 hammer2_chain_lookup_done(parent);
1285         if (ochain)
1286                 hammer2_chain_drop(ochain);
1287
1288         return error;
1289 }
1290
1291 /*
1292  * Given an exclusively locked inode and chain we consolidate its chain
1293  * for hardlink creation, adding (nlinks) to the file's link count and
1294  * potentially relocating the inode to a directory common to ip->pip and tdip.
1295  *
1296  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1297  * and returning a new locked chain.
1298  *
1299  * NOTE!  This function will also replace ip->chain.
1300  */
1301 int
1302 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1303                              hammer2_chain_t **chainp,
1304                              hammer2_inode_t *tdip, int nlinks)
1305 {
1306         hammer2_inode_data_t *ipdata;
1307         hammer2_inode_t *fdip;
1308         hammer2_inode_t *cdip;
1309         hammer2_chain_t *chain;
1310         hammer2_chain_t *nchain;
1311         int error;
1312
1313         chain = *chainp;
1314         if (nlinks == 0 &&                      /* no hardlink needed */
1315             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1316                 return (0);
1317         }
1318         if (hammer2_hardlink_enable < 0) {      /* fake hardlinks */
1319                 return (0);
1320         }
1321
1322         if (hammer2_hardlink_enable == 0) {     /* disallow hardlinks */
1323                 hammer2_chain_unlock(chain);
1324                 *chainp = NULL;
1325                 return (ENOTSUP);
1326         }
1327
1328         /*
1329          * cdip will be returned with a ref, but not locked.
1330          */
1331         fdip = ip->pip;
1332         cdip = hammer2_inode_common_parent(fdip, tdip);
1333
1334         /*
1335          * If no change in the hardlink's target directory is required and
1336          * this is already a hardlink target, all we need to do is adjust
1337          * the link count.
1338          *
1339          * XXX The common parent is a big wiggly due to duplication from
1340          *     renames.  Compare the core (RBTREE) pointer instead of the
1341          *     ip's.
1342          */
1343         if (cdip == fdip &&
1344             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1345                 if (nlinks) {
1346                         hammer2_chain_modify(trans, &chain, 0);
1347                         chain->data->ipdata.nlinks += nlinks;
1348                 }
1349                 error = 0;
1350                 goto done;
1351         }
1352
1353
1354         /*
1355          * chain is the real inode.  If it's visible we have to convert it
1356          * to a hardlink pointer.  If it is not visible then it is already
1357          * a hardlink target and only needs to be deleted.
1358          */
1359         KKASSERT((chain->flags & HAMMER2_CHAIN_DELETED) == 0);
1360         KKASSERT(chain->data->ipdata.type != HAMMER2_OBJTYPE_HARDLINK);
1361         if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1362                 /*
1363                  * We are going to duplicate chain later, causing its
1364                  * media block to be shifted to the duplicate.  Even though
1365                  * we are delete-duplicating nchain here it might decide not
1366                  * to reallocate the block.  Set FORCECOW to force it to.
1367                  */
1368                 nchain = chain;
1369                 hammer2_chain_lock(nchain, HAMMER2_RESOLVE_ALWAYS);
1370                 atomic_set_int(&nchain->flags, HAMMER2_CHAIN_FORCECOW);
1371                 hammer2_chain_delete_duplicate(trans, &nchain,
1372                                                HAMMER2_DELDUP_RECORE);
1373                 KKASSERT((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0);
1374
1375                 ipdata = &nchain->data->ipdata;
1376                 ipdata->target_type = ipdata->type;
1377                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1378                 ipdata->uflags = 0;
1379                 ipdata->rmajor = 0;
1380                 ipdata->rminor = 0;
1381                 ipdata->ctime = 0;
1382                 ipdata->mtime = 0;
1383                 ipdata->atime = 0;
1384                 ipdata->btime = 0;
1385                 bzero(&ipdata->uid, sizeof(ipdata->uid));
1386                 bzero(&ipdata->gid, sizeof(ipdata->gid));
1387                 ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1388                 ipdata->cap_flags = 0;
1389                 ipdata->mode = 0;
1390                 ipdata->size = 0;
1391                 ipdata->nlinks = 1;
1392                 ipdata->iparent = 0;    /* XXX */
1393                 ipdata->pfs_type = 0;
1394                 ipdata->pfs_inum = 0;
1395                 bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1396                 bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1397                 ipdata->data_quota = 0;
1398                 ipdata->data_count = 0;
1399                 ipdata->inode_quota = 0;
1400                 ipdata->inode_count = 0;
1401                 ipdata->attr_tid = 0;
1402                 ipdata->dirent_tid = 0;
1403                 bzero(&ipdata->u, sizeof(ipdata->u));
1404                 /* XXX transaction ids */
1405         } else {
1406                 hammer2_chain_delete(trans, chain, 0);
1407                 nchain = NULL;
1408         }
1409
1410         /*
1411          * chain represents the hardlink target and is now flagged deleted.
1412          * duplicate it to the parent directory and adjust nlinks.
1413          *
1414          * WARNING! The shiftup() call can cause nchain to be moved into
1415          *          an indirect block, and our nchain will wind up pointing
1416          *          to the older/original version.
1417          */
1418         KKASSERT(chain->flags & HAMMER2_CHAIN_DELETED);
1419         hammer2_hardlink_shiftup(trans, &chain, cdip, nlinks, &error);
1420
1421         if (error == 0)
1422                 hammer2_inode_repoint(ip, cdip, chain);
1423
1424         /*
1425          * Unlock the original chain last as the lock blocked races against
1426          * the creation of the new hardlink target.
1427          */
1428         if (nchain)
1429                 hammer2_chain_unlock(nchain);
1430
1431 done:
1432         /*
1433          * Cleanup, chain/nchain already dealt with.
1434          */
1435         *chainp = chain;
1436         hammer2_inode_drop(cdip);
1437
1438         return (error);
1439 }
1440
1441 /*
1442  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1443  * inode while (*chainp) points to the resolved (hidden hardlink
1444  * target) inode.  In this situation when nlinks is 1 we wish to
1445  * deconsolidate the hardlink, moving it back to the directory that now
1446  * represents the only remaining link.
1447  */
1448 int
1449 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1450                                hammer2_inode_t *dip,
1451                                hammer2_chain_t **chainp,
1452                                hammer2_chain_t **ochainp)
1453 {
1454         if (*ochainp == NULL)
1455                 return (0);
1456         /* XXX */
1457         return (0);
1458 }
1459
1460 /*
1461  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1462  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1463  * the *chainp and return a new locked *chainp representing the file target
1464  * (the original *chainp will be unlocked).
1465  *
1466  * When a match is found the chain representing the original HARDLINK
1467  * will be returned in *ochainp with a ref, but not locked.
1468  *
1469  * When no match is found *chainp is set to NULL and EIO is returned.
1470  * (*ochainp) will still be set to the original chain with a ref but not
1471  * locked.
1472  */
1473 int
1474 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1475                       hammer2_chain_t **ochainp)
1476 {
1477         hammer2_chain_t *chain = *chainp;
1478         hammer2_chain_t *parent;
1479         hammer2_inode_t *ip;
1480         hammer2_inode_t *pip;
1481         hammer2_key_t key_dummy;
1482         hammer2_key_t lhc;
1483         int cache_index = -1;
1484
1485         pip = dip;
1486         hammer2_inode_ref(pip);         /* for loop */
1487         hammer2_chain_ref(chain);       /* for (*ochainp) */
1488         *ochainp = chain;
1489
1490         /*
1491          * Locate the hardlink.  pip is referenced and not locked,
1492          * ipp.
1493          *
1494          * chain is reused.
1495          */
1496         lhc = chain->data->ipdata.inum;
1497         hammer2_chain_unlock(chain);
1498         chain = NULL;
1499
1500         while ((ip = pip) != NULL) {
1501                 parent = hammer2_inode_lock_ex(ip);
1502                 hammer2_inode_drop(ip);                 /* loop */
1503                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1504                 chain = hammer2_chain_lookup(&parent, &key_dummy,
1505                                              lhc, lhc, &cache_index, 0);
1506                 hammer2_chain_lookup_done(parent);      /* discard parent */
1507                 if (chain)
1508                         break;
1509                 pip = ip->pip;          /* safe, ip held locked */
1510                 if (pip)
1511                         hammer2_inode_ref(pip);         /* loop */
1512                 hammer2_inode_unlock_ex(ip, NULL);
1513         }
1514
1515         /*
1516          * chain is locked, ip is locked.  Unlock ip, return the locked
1517          * chain.  *ipp is already set w/a ref count and not locked.
1518          *
1519          * (parent is already unlocked).
1520          */
1521         if (ip)
1522                 hammer2_inode_unlock_ex(ip, NULL);
1523         *chainp = chain;
1524         if (chain) {
1525                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1526                 /* already locked */
1527                 return (0);
1528         } else {
1529                 return (EIO);
1530         }
1531 }
1532
1533 /*
1534  * Find the directory common to both fdip and tdip, hold and return
1535  * its inode.
1536  */
1537 hammer2_inode_t *
1538 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1539 {
1540         hammer2_inode_t *scan1;
1541         hammer2_inode_t *scan2;
1542
1543         /*
1544          * We used to have a depth field but it complicated matters too
1545          * much for directory renames.  So now its ugly.  Check for
1546          * simple cases before giving up and doing it the expensive way.
1547          *
1548          * XXX need a bottom-up topology stability lock
1549          */
1550         if (fdip == tdip || fdip == tdip->pip) {
1551                 hammer2_inode_ref(fdip);
1552                 return(fdip);
1553         }
1554         if (fdip->pip == tdip) {
1555                 hammer2_inode_ref(tdip);
1556                 return(tdip);
1557         }
1558
1559         /*
1560          * XXX not MPSAFE
1561          */
1562         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1563                 scan2 = tdip;
1564                 while (scan2->pmp == tdip->pmp) {
1565                         if (scan1 == scan2) {
1566                                 hammer2_inode_ref(scan1);
1567                                 return(scan1);
1568                         }
1569                         scan2 = scan2->pip;
1570                         if (scan2 == NULL)
1571                                 break;
1572                 }
1573         }
1574         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1575               fdip, tdip);
1576         /* NOT REACHED */
1577         return(NULL);
1578 }
1579
1580 /*
1581  * Synchronize the inode's frontend state with the chain state prior
1582  * to any explicit flush of the inode or any strategy write call.
1583  *
1584  * Called with a locked inode.
1585  */
1586 void
1587 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1588                     hammer2_chain_t **chainp)
1589 {
1590         hammer2_inode_data_t *ipdata;
1591         hammer2_chain_t *parent;
1592         hammer2_chain_t *chain;
1593         hammer2_key_t lbase;
1594         hammer2_key_t key_next;
1595         int cache_index;
1596
1597         ipdata = &ip->chain->data->ipdata;
1598
1599         if (ip->flags & HAMMER2_INODE_MTIME) {
1600                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1601                 atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1602                 ipdata->mtime = ip->mtime;
1603         }
1604         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1605                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1606                 ipdata->size = ip->size;
1607                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1608
1609                 /*
1610                  * We must delete any chains beyond the EOF.  The chain
1611                  * straddling the EOF will be pending in the bioq.
1612                  */
1613                 lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1614                         ~HAMMER2_PBUFMASK64;
1615                 parent = hammer2_chain_lookup_init(ip->chain, 0);
1616                 chain = hammer2_chain_lookup(&parent, &key_next,
1617                                              lbase, (hammer2_key_t)-1,
1618                                              &cache_index,
1619                                              HAMMER2_LOOKUP_NODATA);
1620                 while (chain) {
1621                         /*
1622                          * Degenerate embedded case, nothing to loop on
1623                          */
1624                         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1625                                 hammer2_chain_unlock(chain);
1626                                 break;
1627                         }
1628                         if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1629                                 hammer2_chain_delete(trans, chain, 0);
1630                         }
1631                         chain = hammer2_chain_next(&parent, chain, &key_next,
1632                                                    key_next, (hammer2_key_t)-1,
1633                                                    &cache_index,
1634                                                    HAMMER2_LOOKUP_NODATA);
1635                 }
1636                 hammer2_chain_lookup_done(parent);
1637         } else
1638         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1639                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1640                 ipdata->size = ip->size;
1641                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1642
1643                 /*
1644                  * When resizing larger we may not have any direct-data
1645                  * available.
1646                  */
1647                 if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1648                     ip->size > HAMMER2_EMBEDDED_BYTES) {
1649                         ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1650                         bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
1651                 }
1652         }
1653 }