ed72877f6d3af6f899560592741baff40eed82d1
[dragonfly.git] / 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 which can cause ip's cached
67  * 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_core_t *core;
79
80         hammer2_inode_ref(ip);
81         ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
82
83         chain = ip->chain;
84         core = chain->core;
85         for (;;) {
86                 if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
87                         spin_lock(&core->cst.spin);
88                         while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
89                                 chain = TAILQ_NEXT(chain, core_entry);
90                         hammer2_chain_ref(chain);
91                         spin_unlock(&core->cst.spin);
92                         hammer2_inode_repoint(ip, NULL, chain);
93                         hammer2_chain_drop(chain);
94                 }
95                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
96                 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
97                         break;
98                 hammer2_chain_unlock(chain);
99         }
100         return (chain);
101 }
102
103 void
104 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
105 {
106         /*
107          * XXX this will catch parent directories too which we don't
108          *     really want.
109          */
110         if (chain)
111                 hammer2_chain_unlock(chain);
112         ccms_thread_unlock(&ip->topo_cst);
113         hammer2_inode_drop(ip);
114 }
115
116 /*
117  * NOTE: We don't combine the inode/chain lock because putting away an
118  *       inode would otherwise confuse multiple lock holders of the inode.
119  *
120  *       Shared locks are especially sensitive to having too many shared
121  *       lock counts (from the same thread) on certain paths which might
122  *       need to upgrade them.  Only one count of a shared lock can be
123  *       upgraded.
124  */
125 hammer2_chain_t *
126 hammer2_inode_lock_sh(hammer2_inode_t *ip)
127 {
128         hammer2_chain_t *chain;
129
130         hammer2_inode_ref(ip);
131         for (;;) {
132                 ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
133
134                 chain = ip->chain;
135                 KKASSERT(chain != NULL);        /* for now */
136                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
137                                           HAMMER2_RESOLVE_SHARED);
138
139                 /*
140                  * Resolve duplication races
141                  */
142                 if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
143                         break;
144                 hammer2_chain_unlock(chain);
145                 ccms_thread_unlock(&ip->topo_cst);
146                 chain = hammer2_inode_lock_ex(ip);
147                 hammer2_inode_unlock_ex(ip, chain);
148         }
149         return (chain);
150 }
151
152 void
153 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
154 {
155         if (chain)
156                 hammer2_chain_unlock(chain);
157         ccms_thread_unlock(&ip->topo_cst);
158         hammer2_inode_drop(ip);
159 }
160
161 ccms_state_t
162 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
163 {
164         return(ccms_thread_lock_temp_release(&ip->topo_cst));
165 }
166
167 void
168 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
169 {
170         ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
171 }
172
173 ccms_state_t
174 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
175 {
176         return(ccms_thread_lock_upgrade(&ip->topo_cst));
177 }
178
179 void
180 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
181 {
182         ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
183 }
184
185 /*
186  * Lookup an inode by inode number
187  */
188 hammer2_inode_t *
189 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
190 {
191         hammer2_inode_t *ip;
192
193         if (pmp) {
194                 spin_lock(&pmp->inum_spin);
195                 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
196                 if (ip)
197                         hammer2_inode_ref(ip);
198                 spin_unlock(&pmp->inum_spin);
199         } else {
200                 ip = NULL;
201         }
202         return(ip);
203 }
204
205 /*
206  * Adding a ref to an inode is only legal if the inode already has at least
207  * one ref.
208  */
209 void
210 hammer2_inode_ref(hammer2_inode_t *ip)
211 {
212         atomic_add_int(&ip->refs, 1);
213 }
214
215 /*
216  * Drop an inode reference, freeing the inode when the last reference goes
217  * away.
218  */
219 void
220 hammer2_inode_drop(hammer2_inode_t *ip)
221 {
222         hammer2_pfsmount_t *pmp;
223         hammer2_inode_t *pip;
224         u_int refs;
225
226         while (ip) {
227                 refs = ip->refs;
228                 cpu_ccfence();
229                 if (refs == 1) {
230                         /*
231                          * Transition to zero, must interlock with
232                          * the inode inumber lookup tree (if applicable).
233                          *
234                          * NOTE: The super-root inode has no pmp.
235                          */
236                         pmp = ip->pmp;
237                         if (pmp)
238                                 spin_lock(&pmp->inum_spin);
239
240                         if (atomic_cmpset_int(&ip->refs, 1, 0)) {
241                                 KKASSERT(ip->topo_cst.count == 0);
242                                 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
243                                         atomic_clear_int(&ip->flags,
244                                                      HAMMER2_INODE_ONRBTREE);
245                                         RB_REMOVE(hammer2_inode_tree,
246                                                   &pmp->inum_tree, ip);
247                                 }
248                                 if (pmp)
249                                         spin_unlock(&pmp->inum_spin);
250
251                                 pip = ip->pip;
252                                 ip->pip = NULL;
253                                 ip->pmp = NULL;
254
255                                 /*
256                                  * Cleaning out ip->chain isn't entirely
257                                  * trivial.
258                                  */
259                                 hammer2_inode_repoint(ip, NULL, NULL);
260
261                                 /*
262                                  * We have to drop pip (if non-NULL) to
263                                  * dispose of our implied reference from
264                                  * ip->pip.  We can simply loop on it.
265                                  */
266                                 if (pmp) {
267                                         KKASSERT((ip->flags &
268                                                   HAMMER2_INODE_SROOT) == 0);
269                                         kfree(ip, pmp->minode);
270                                         atomic_add_long(&pmp->inmem_inodes, -1);
271                                 } else {
272                                         KKASSERT(ip->flags &
273                                                  HAMMER2_INODE_SROOT);
274                                         kfree(ip, M_HAMMER2);
275                                 }
276                                 ip = pip;
277                                 /* continue with pip (can be NULL) */
278                         } else {
279                                 if (pmp)
280                                         spin_unlock(&ip->pmp->inum_spin);
281                         }
282                 } else {
283                         /*
284                          * Non zero transition
285                          */
286                         if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
287                                 break;
288                 }
289         }
290 }
291
292 /*
293  * Get the vnode associated with the given inode, allocating the vnode if
294  * necessary.  The vnode will be returned exclusively locked.
295  *
296  * The caller must lock the inode (shared or exclusive).
297  *
298  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
299  * races.
300  */
301 struct vnode *
302 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
303 {
304         hammer2_inode_data_t *ipdata;
305         hammer2_pfsmount_t *pmp;
306         struct vnode *vp;
307         ccms_state_t ostate;
308
309         pmp = ip->pmp;
310         KKASSERT(pmp != NULL);
311         *errorp = 0;
312         ipdata = &ip->chain->data->ipdata;
313
314         for (;;) {
315                 /*
316                  * Attempt to reuse an existing vnode assignment.  It is
317                  * possible to race a reclaim so the vget() may fail.  The
318                  * inode must be unlocked during the vget() to avoid a
319                  * deadlock against a reclaim.
320                  */
321                 vp = ip->vp;
322                 if (vp) {
323                         /*
324                          * Inode must be unlocked during the vget() to avoid
325                          * possible deadlocks, but leave the ip ref intact.
326                          *
327                          * vnode is held to prevent destruction during the
328                          * vget().  The vget() can still fail if we lost
329                          * a reclaim race on the vnode.
330                          */
331                         vhold_interlocked(vp);
332                         ostate = hammer2_inode_lock_temp_release(ip);
333                         if (vget(vp, LK_EXCLUSIVE)) {
334                                 vdrop(vp);
335                                 hammer2_inode_lock_temp_restore(ip, ostate);
336                                 continue;
337                         }
338                         hammer2_inode_lock_temp_restore(ip, ostate);
339                         vdrop(vp);
340                         /* vp still locked and ref from vget */
341                         if (ip->vp != vp) {
342                                 kprintf("hammer2: igetv race %p/%p\n",
343                                         ip->vp, vp);
344                                 vput(vp);
345                                 continue;
346                         }
347                         *errorp = 0;
348                         break;
349                 }
350
351                 /*
352                  * No vnode exists, allocate a new vnode.  Beware of
353                  * allocation races.  This function will return an
354                  * exclusively locked and referenced vnode.
355                  */
356                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
357                 if (*errorp) {
358                         kprintf("hammer2: igetv getnewvnode failed %d\n",
359                                 *errorp);
360                         vp = NULL;
361                         break;
362                 }
363
364                 /*
365                  * Lock the inode and check for an allocation race.
366                  */
367                 ostate = hammer2_inode_lock_upgrade(ip);
368                 if (ip->vp != NULL) {
369                         vp->v_type = VBAD;
370                         vx_put(vp);
371                         hammer2_inode_lock_downgrade(ip, ostate);
372                         continue;
373                 }
374
375                 switch (ipdata->type) {
376                 case HAMMER2_OBJTYPE_DIRECTORY:
377                         vp->v_type = VDIR;
378                         break;
379                 case HAMMER2_OBJTYPE_REGFILE:
380                         vp->v_type = VREG;
381                         vinitvmio(vp, ipdata->size,
382                                   HAMMER2_LBUFSIZE,
383                                   (int)ipdata->size & HAMMER2_LBUFMASK);
384                         break;
385                 case HAMMER2_OBJTYPE_SOFTLINK:
386                         /*
387                          * XXX for now we are using the generic file_read
388                          * and file_write code so we need a buffer cache
389                          * association.
390                          */
391                         vp->v_type = VLNK;
392                         vinitvmio(vp, ipdata->size,
393                                   HAMMER2_LBUFSIZE,
394                                   (int)ipdata->size & HAMMER2_LBUFMASK);
395                         break;
396                 /* XXX FIFO */
397                 default:
398                         panic("hammer2: unhandled objtype %d", ipdata->type);
399                         break;
400                 }
401
402                 if (ip == pmp->iroot)
403                         vsetflags(vp, VROOT);
404
405                 vp->v_data = ip;
406                 ip->vp = vp;
407                 hammer2_inode_ref(ip);          /* vp association */
408                 hammer2_inode_lock_downgrade(ip, ostate);
409                 break;
410         }
411
412         /*
413          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
414          */
415         if (hammer2_debug & 0x0002) {
416                 kprintf("igetv vp %p refs %d aux %d\n",
417                         vp, vp->v_sysref.refcnt, vp->v_auxrefs);
418         }
419         return (vp);
420 }
421
422 /*
423  * The passed-in chain must be locked and the returned inode will also be
424  * locked.  This routine typically locates or allocates the inode, assigns
425  * ip->chain (adding a ref to chain if necessary), and returns the inode.
426  *
427  * The hammer2_inode structure regulates the interface between the high level
428  * kernel VNOPS API and the filesystem backend (the chains).
429  *
430  * WARNING!  This routine sucks up the chain's lock (makes it part of the
431  *           inode lock from the point of view of the inode lock API),
432  *           so callers need to be careful.
433  *
434  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
435  *           is allowed to pass pmp == NULL and dip == NULL for sroot.
436  */
437 hammer2_inode_t *
438 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
439                   hammer2_chain_t *chain)
440 {
441         hammer2_inode_t *nip;
442
443         KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
444
445         /*
446          * Interlocked lookup/ref of the inode.  This code is only needed
447          * when looking up inodes with nlinks != 0 (TODO: optimize out
448          * otherwise and test for duplicates).
449          */
450 again:
451         for (;;) {
452                 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
453                 if (nip == NULL)
454                         break;
455                 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
456                 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
457                         ccms_thread_unlock(&nip->topo_cst);
458                         hammer2_inode_drop(nip);
459                         continue;
460                 }
461                 if (nip->chain != chain)
462                         hammer2_inode_repoint(nip, NULL, chain);
463
464                 /*
465                  * Consolidated nip/nip->chain is locked (chain locked
466                  * by caller).
467                  */
468                 return nip;
469         }
470
471         /*
472          * We couldn't find the inode number, create a new inode.
473          */
474         if (pmp) {
475                 nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
476                 atomic_add_long(&pmp->inmem_inodes, 1);
477                 hammer2_chain_memory_wakeup(pmp);
478         } else {
479                 nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
480                 nip->flags = HAMMER2_INODE_SROOT;
481         }
482         nip->inum = chain->data->ipdata.inum;
483         nip->size = chain->data->ipdata.size;
484         nip->mtime = chain->data->ipdata.mtime;
485         hammer2_inode_repoint(nip, NULL, chain);
486         nip->pip = dip;                         /* can be NULL */
487         if (dip)
488                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
489
490         nip->pmp = pmp;
491
492         /*
493          * ref and lock on nip gives it state compatible to after a
494          * hammer2_inode_lock_ex() call.
495          */
496         nip->refs = 1;
497         ccms_cst_init(&nip->topo_cst, &nip->chain);
498         ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
499         /* combination of thread lock and chain lock == inode lock */
500
501         /*
502          * Attempt to add the inode.  If it fails we raced another inode
503          * get.  Undo all the work and try again.
504          */
505         if (pmp) {
506                 spin_lock(&pmp->inum_spin);
507                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
508                         spin_unlock(&pmp->inum_spin);
509                         ccms_thread_unlock(&nip->topo_cst);
510                         hammer2_inode_drop(nip);
511                         goto again;
512                 }
513                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
514                 spin_unlock(&pmp->inum_spin);
515         }
516
517         return (nip);
518 }
519
520 /*
521  * Create a new inode in the specified directory using the vattr to
522  * figure out the type of inode.
523  *
524  * If no error occurs the new inode with its chain locked is returned in
525  * *nipp, otherwise an error is returned and *nipp is set to NULL.
526  *
527  * If vap and/or cred are NULL the related fields are not set and the
528  * inode type defaults to a directory.  This is used when creating PFSs
529  * under the super-root, so the inode number is set to 1 in this case.
530  *
531  * dip is not locked on entry.
532  */
533 hammer2_inode_t *
534 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
535                      struct vattr *vap, struct ucred *cred,
536                      const uint8_t *name, size_t name_len,
537                      hammer2_chain_t **chainp, int *errorp)
538 {
539         hammer2_inode_data_t *dipdata;
540         hammer2_inode_data_t *nipdata;
541         hammer2_chain_t *chain;
542         hammer2_chain_t *parent;
543         hammer2_inode_t *nip;
544         hammer2_key_t key_dummy;
545         hammer2_key_t lhc;
546         int error;
547         uid_t xuid;
548         uuid_t dip_uid;
549         uuid_t dip_gid;
550         uint32_t dip_mode;
551         uint8_t dip_algo;
552         int cache_index = -1;
553
554         lhc = hammer2_dirhash(name, name_len);
555         *errorp = 0;
556
557         /*
558          * Locate the inode or indirect block to create the new
559          * entry in.  At the same time check for key collisions
560          * and iterate until we don't get one.
561          *
562          * NOTE: hidden inodes do not have iterators.
563          */
564 retry:
565         parent = hammer2_inode_lock_ex(dip);
566         dipdata = &dip->chain->data->ipdata;
567         dip_uid = dipdata->uid;
568         dip_gid = dipdata->gid;
569         dip_mode = dipdata->mode;
570         dip_algo = dipdata->comp_algo;
571
572         error = 0;
573         while (error == 0) {
574                 chain = hammer2_chain_lookup(&parent, &key_dummy,
575                                              lhc, lhc, &cache_index, 0);
576                 if (chain == NULL)
577                         break;
578                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
579                         error = ENOSPC;
580                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
581                         error = ENOSPC;
582                 hammer2_chain_unlock(chain);
583                 chain = NULL;
584                 ++lhc;
585         }
586
587         if (error == 0) {
588                 error = hammer2_chain_create(trans, &parent, &chain,
589                                              lhc, 0,
590                                              HAMMER2_BREF_TYPE_INODE,
591                                              HAMMER2_INODE_BYTES);
592         }
593
594         /*
595          * Cleanup and handle retries.
596          */
597         if (error == EAGAIN) {
598                 hammer2_chain_ref(parent);
599                 hammer2_inode_unlock_ex(dip, parent);
600                 hammer2_chain_wait(parent);
601                 hammer2_chain_drop(parent);
602                 goto retry;
603         }
604         hammer2_inode_unlock_ex(dip, parent);
605
606         if (error) {
607                 KKASSERT(chain == NULL);
608                 *errorp = error;
609                 return (NULL);
610         }
611
612         /*
613          * Set up the new inode.
614          *
615          * NOTE: *_get() integrates chain's lock into the inode lock.
616          *
617          * NOTE: Only one new inode can currently be created per
618          *       transaction.  If the need arises we can adjust
619          *       hammer2_trans_init() to allow more.
620          *
621          * NOTE: nipdata will have chain's blockset data.
622          */
623         chain->data->ipdata.inum = trans->sync_tid;
624         nip = hammer2_inode_get(dip->pmp, dip, chain);
625         nipdata = &chain->data->ipdata;
626
627         if (vap) {
628                 KKASSERT(trans->inodes_created == 0);
629                 nipdata->type = hammer2_get_obj_type(vap->va_type);
630                 nipdata->inum = trans->sync_tid;
631                 ++trans->inodes_created;
632         } else {
633                 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
634                 nipdata->inum = 1;
635         }
636         
637         /* Inherit parent's inode compression mode. */
638         nip->comp_heuristic = 0;
639         nipdata->comp_algo = dip_algo;
640         nipdata->version = HAMMER2_INODE_VERSION_ONE;
641         hammer2_update_time(&nipdata->ctime);
642         nipdata->mtime = nipdata->ctime;
643         if (vap)
644                 nipdata->mode = vap->va_mode;
645         nipdata->nlinks = 1;
646         if (vap) {
647                 if (dip) {
648                         xuid = hammer2_to_unix_xid(&dip_uid);
649                         xuid = vop_helper_create_uid(dip->pmp->mp,
650                                                      dip_mode,
651                                                      xuid,
652                                                      cred,
653                                                      &vap->va_mode);
654                 } else {
655                         xuid = 0;
656                 }
657                 if (vap->va_vaflags & VA_UID_UUID_VALID)
658                         nipdata->uid = vap->va_uid_uuid;
659                 else if (vap->va_uid != (uid_t)VNOVAL)
660                         hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
661                 else
662                         hammer2_guid_to_uuid(&nipdata->uid, xuid);
663
664                 if (vap->va_vaflags & VA_GID_UUID_VALID)
665                         nipdata->gid = vap->va_gid_uuid;
666                 else if (vap->va_gid != (gid_t)VNOVAL)
667                         hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
668                 else if (dip)
669                         nipdata->gid = dip_gid;
670         }
671
672         /*
673          * Regular files and softlinks allow a small amount of data to be
674          * directly embedded in the inode.  This flag will be cleared if
675          * the size is extended past the embedded limit.
676          */
677         if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
678             nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
679                 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
680         }
681
682         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
683         bcopy(name, nipdata->filename, name_len);
684         nipdata->name_key = lhc;
685         nipdata->name_len = name_len;
686         *chainp = chain;
687
688         return (nip);
689 }
690
691 /*
692  * chain may have been moved around by the create.
693  */
694 static
695 void
696 hammer2_chain_refactor(hammer2_chain_t **chainp)
697 {
698         hammer2_chain_t *chain = *chainp;
699         hammer2_chain_core_t *core;
700
701         core = chain->core;
702         while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
703                 spin_lock(&core->cst.spin);
704                 chain = TAILQ_NEXT(chain, core_entry);
705                 while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
706                         chain = TAILQ_NEXT(chain, core_entry);
707                 hammer2_chain_ref(chain);
708                 spin_unlock(&core->cst.spin);
709                 KKASSERT(chain->core == core);
710
711                 hammer2_chain_unlock(*chainp);
712                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
713                                           HAMMER2_RESOLVE_NOREF); /* eat ref */
714                 *chainp = chain;
715         }
716 }
717
718 /*
719  * ochain represents the target file inode.  We need to move it to the
720  * specified common parent directory (dip) and rename it to a special
721  * invisible "0xINODENUMBER" filename.
722  *
723  * We use chain_duplicate and duplicate ochain at the new location,
724  * renaming it appropriately.  We create a temporary chain and
725  * then delete it to placemark where the duplicate will go.  Both of
726  * these use the inode number for (lhc) (the key), generating the
727  * invisible filename.
728  */
729 static
730 hammer2_chain_t *
731 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
732                         hammer2_inode_t *dip, int *errorp)
733 {
734         hammer2_inode_data_t *nipdata;
735         hammer2_chain_t *parent;
736         hammer2_chain_t *ochain;
737         hammer2_chain_t *nchain;
738         hammer2_chain_t *tmp;
739         hammer2_key_t key_dummy;
740         hammer2_key_t lhc;
741         hammer2_blockref_t bref;
742         int cache_index = -1;
743
744         ochain = *ochainp;
745         *errorp = 0;
746         lhc = ochain->data->ipdata.inum;
747         KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
748
749         /*
750          * Locate the inode or indirect block to create the new
751          * entry in.  lhc represents the inode number so there is
752          * no collision iteration.
753          *
754          * There should be no key collisions with invisible inode keys.
755          *
756          * WARNING! Must use inode_lock_ex() on dip to handle a stale
757          *          dip->chain cache.
758          */
759 retry:
760         parent = hammer2_inode_lock_ex(dip);
761         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
762         nchain = hammer2_chain_lookup(&parent, &key_dummy,
763                                       lhc, lhc, &cache_index, 0);
764         if (nchain) {
765                 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
766                         nchain, parent, dip, dip->chain);
767                 hammer2_chain_unlock(nchain);
768                 nchain = NULL;
769                 *errorp = ENOSPC;
770 #if 1
771                 Debugger("X3");
772 #endif
773         }
774
775         /*
776          * Create entry in common parent directory using the seek position
777          * calculated above.
778          */
779         if (*errorp == 0) {
780                 KKASSERT(nchain == NULL);
781                 *errorp = hammer2_chain_create(trans, &parent, &nchain,
782                                                lhc, 0,
783                                                HAMMER2_BREF_TYPE_INODE,/* n/a */
784                                                HAMMER2_INODE_BYTES);   /* n/a */
785                 hammer2_chain_refactor(&ochain);
786                 *ochainp = ochain;
787         }
788
789         /*
790          * Cleanup and handle retries.
791          */
792         if (*errorp == EAGAIN) {
793                 hammer2_chain_ref(parent);
794                 /* hammer2_chain_lookup_done(parent); */
795                 hammer2_inode_unlock_ex(dip, parent);
796                 hammer2_chain_wait(parent);
797                 hammer2_chain_drop(parent);
798                 goto retry;
799         }
800
801         /*
802          * Handle the error case
803          */
804         if (*errorp) {
805                 KKASSERT(nchain == NULL);
806                 hammer2_inode_unlock_ex(dip, parent);
807                 /*hammer2_chain_lookup_done(parent);*/
808                 return (NULL);
809         }
810
811         /*
812          * Use chain as a placeholder for (lhc), delete it and replace
813          * it with our duplication.
814          *
815          * Gain a second lock on ochain for the duplication function to
816          * unlock, maintain the caller's original lock across the call.
817          *
818          * This is a bit messy.
819          */
820         hammer2_chain_delete(trans, nchain, HAMMER2_DELETE_WILLDUP);
821         hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
822         tmp = ochain;
823         bref = tmp->bref;
824         bref.key = lhc;                 /* invisible dir entry key */
825         bref.keybits = 0;
826         hammer2_chain_duplicate(trans, parent, &tmp, &bref, 0);
827         hammer2_inode_unlock_ex(dip, parent);
828         /*hammer2_chain_lookup_done(parent);*/
829         hammer2_chain_unlock(nchain);   /* no longer needed */
830
831         /*
832          * Now set chain to our duplicate and modify it appropriately.
833          *
834          * Directory entries are inodes but this is a hidden hardlink
835          * target.  The name isn't used but to ease debugging give it
836          * a name after its inode number.
837          */
838         nchain = tmp;
839         tmp = NULL;     /* safety */
840
841         hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
842         nipdata = &nchain->data->ipdata;
843         ksnprintf(nipdata->filename, sizeof(nipdata->filename),
844                   "0x%016jx", (intmax_t)nipdata->inum);
845         nipdata->name_len = strlen(nipdata->filename);
846         nipdata->name_key = lhc;
847
848         return (nchain);
849 }
850
851 /*
852  * Connect the target inode represented by (*chainp) to the media topology
853  * at (dip, name, len).
854  *
855  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
856  * entry instead of connecting (*chainp).
857  *
858  * If hlink is FALSE this function uses chain_duplicate() to make a copy
859  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
860  * by the caller in this case (e.g. rename).
861  */
862 int
863 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
864                       hammer2_inode_t *dip, hammer2_chain_t **chainp,
865                       const uint8_t *name, size_t name_len)
866 {
867         hammer2_inode_data_t *ipdata;
868         hammer2_chain_t *nchain;
869         hammer2_chain_t *parent;
870         hammer2_chain_t *ochain;
871         hammer2_key_t key_dummy;
872         hammer2_key_t lhc;
873         int cache_index = -1;
874         int error;
875
876         ochain = *chainp;
877
878         /*
879          * Since ochain is either disconnected from the topology or represents
880          * a hardlink terminus which is always a parent of or equal to dip,
881          * we should be able to safely lock dip->chain for our setup.
882          *
883          * WARNING! Must use inode_lock_ex() on dip to handle a stale
884          *          dip->chain cache.
885          */
886         parent = hammer2_inode_lock_ex(dip);
887         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
888
889         lhc = hammer2_dirhash(name, name_len);
890
891         /*
892          * Locate the inode or indirect block to create the new
893          * entry in.  At the same time check for key collisions
894          * and iterate until we don't get one.
895          */
896         error = 0;
897         while (error == 0) {
898                 nchain = hammer2_chain_lookup(&parent, &key_dummy,
899                                               lhc, lhc, &cache_index, 0);
900                 if (nchain == NULL)
901                         break;
902                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
903                         error = ENOSPC;
904                 hammer2_chain_unlock(nchain);
905                 nchain = NULL;
906                 ++lhc;
907         }
908
909         if (error == 0) {
910                 if (hlink) {
911                         /*
912                          * Hardlink pointer needed, create totally fresh
913                          * directory entry.
914                          */
915                         KKASSERT(nchain == NULL);
916                         error = hammer2_chain_create(trans, &parent, &nchain,
917                                                      lhc, 0,
918                                                      HAMMER2_BREF_TYPE_INODE,
919                                                      HAMMER2_INODE_BYTES);
920                         hammer2_chain_refactor(&ochain);
921                 } else {
922                         /*
923                          * Reconnect the original chain and rename.  Use
924                          * chain_duplicate().  The caller will likely delete
925                          * or has already deleted the original chain in
926                          * this case.
927                          *
928                          * NOTE: chain_duplicate() generates a new chain
929                          *       with CHAIN_DELETED cleared (ochain typically
930                          *       has it set from the file unlink).
931                          */
932                         nchain = ochain;
933                         ochain = NULL;
934                         hammer2_chain_duplicate(trans, NULL, &nchain, NULL, 0);
935                         error = hammer2_chain_create(trans, &parent, &nchain,
936                                                      lhc, 0,
937                                                      HAMMER2_BREF_TYPE_INODE,
938                                                      HAMMER2_INODE_BYTES);
939                 }
940         }
941
942         /*
943          * Unlock stuff.
944          */
945         KKASSERT(error != EAGAIN);
946         hammer2_inode_unlock_ex(dip, parent);
947         /*hammer2_chain_lookup_done(parent);*/
948         parent = NULL;
949
950         /*
951          * nchain should be NULL on error, leave ochain (== *chainp) alone.
952          */
953         if (error) {
954                 KKASSERT(nchain == NULL);
955                 return (error);
956         }
957
958         /*
959          * Directory entries are inodes so if the name has changed we have
960          * to update the inode.
961          *
962          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
963          * chain, the caller will access the hardlink via the actual hardlink
964          * target file and not the hardlink pointer entry, so we must still
965          * return ochain.
966          */
967         if (hlink && hammer2_hardlink_enable >= 0) {
968                 /*
969                  * Create the HARDLINK pointer.  oip represents the hardlink
970                  * target in this situation.
971                  *
972                  * We will return ochain (the hardlink target).
973                  */
974                 hammer2_chain_modify(trans, &nchain,
975                                      HAMMER2_MODIFY_ASSERTNOCOPY);
976                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
977                 ipdata = &nchain->data->ipdata;
978                 bcopy(name, ipdata->filename, name_len);
979                 ipdata->name_key = lhc;
980                 ipdata->name_len = name_len;
981                 ipdata->target_type = ochain->data->ipdata.type;
982                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
983                 ipdata->inum = ochain->data->ipdata.inum;
984                 ipdata->nlinks = 1;
985                 hammer2_chain_unlock(nchain);
986                 nchain = ochain;
987                 ochain = NULL;
988         } else if (hlink && hammer2_hardlink_enable < 0) {
989                 /*
990                  * Create a snapshot (hardlink fake mode for debugging).
991                  * (ochain already flushed above so we can just copy the
992                  * bref XXX).
993                  *
994                  * Since this is a snapshot we return nchain in the fake
995                  * hardlink case.
996                  */
997                 hammer2_chain_modify(trans, &nchain,
998                                      HAMMER2_MODIFY_ASSERTNOCOPY);
999                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1000                 ipdata = &nchain->data->ipdata;
1001                 *ipdata = ochain->data->ipdata;
1002                 bcopy(name, ipdata->filename, name_len);
1003                 ipdata->name_key = lhc;
1004                 ipdata->name_len = name_len;
1005                 kprintf("created fake hardlink %*.*s\n",
1006                         (int)name_len, (int)name_len, name);
1007         } else {
1008                 /*
1009                  * nchain is a duplicate of ochain at the new location.
1010                  * We must fixup the name stored in oip.  The bref key
1011                  * has already been set up.
1012                  */
1013                 hammer2_chain_modify(trans, &nchain,
1014                                      HAMMER2_MODIFY_ASSERTNOCOPY);
1015                 ipdata = &nchain->data->ipdata;
1016
1017                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1018                 bcopy(name, ipdata->filename, name_len);
1019                 ipdata->name_key = lhc;
1020                 ipdata->name_len = name_len;
1021                 ipdata->nlinks = 1;
1022         }
1023
1024         /*
1025          * We are replacing ochain with nchain, unlock ochain.  In the
1026          * case where ochain is left unchanged the code above sets
1027          * nchain to ochain and ochain to NULL, resulting in a NOP here.
1028          */
1029         if (ochain)
1030                 hammer2_chain_unlock(ochain);
1031         *chainp = nchain;
1032
1033         return (0);
1034 }
1035
1036 /*
1037  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1038  * locked.
1039  *
1040  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1041  * and nchain is ref'd.
1042  */
1043 void
1044 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1045                       hammer2_chain_t *nchain)
1046 {
1047         hammer2_chain_t *ochain;
1048         hammer2_inode_t *opip;
1049
1050         /*
1051          * Repoint ip->chain if requested.
1052          */
1053         ochain = ip->chain;
1054         ip->chain = nchain;
1055         if (nchain)
1056                 hammer2_chain_ref(nchain);
1057         if (ochain)
1058                 hammer2_chain_drop(ochain);
1059
1060         /*
1061          * Repoint ip->pip if requested (non-NULL pip).
1062          */
1063         if (pip && ip->pip != pip) {
1064                 opip = ip->pip;
1065                 hammer2_inode_ref(pip);
1066                 ip->pip = pip;
1067                 if (opip)
1068                         hammer2_inode_drop(opip);
1069         }
1070 }
1071
1072 /*
1073  * Unlink the file from the specified directory inode.  The directory inode
1074  * does not need to be locked.
1075  *
1076  * isdir determines whether a directory/non-directory check should be made.
1077  * No check is made if isdir is set to -1.
1078  *
1079  * NOTE!  The underlying file can still be active with open descriptors
1080  *        or if the chain is being manually held (e.g. for rename).
1081  *
1082  *        The caller is responsible for fixing up ip->chain if e.g. a
1083  *        rename occurs (see chain_duplicate()).
1084  */
1085 int
1086 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1087                     const uint8_t *name, size_t name_len,
1088                     int isdir, int *hlinkp)
1089 {
1090         hammer2_inode_data_t *ipdata;
1091         hammer2_chain_t *parent;
1092         hammer2_chain_t *ochain;
1093         hammer2_chain_t *chain;
1094         hammer2_chain_t *dparent;
1095         hammer2_chain_t *dchain;
1096         hammer2_key_t key_dummy;
1097         hammer2_key_t key_next;
1098         hammer2_key_t lhc;
1099         int error;
1100         int cache_index = -1;
1101         uint8_t type;
1102
1103         error = 0;
1104         ochain = NULL;
1105         lhc = hammer2_dirhash(name, name_len);
1106
1107         /*
1108          * Search for the filename in the directory
1109          */
1110         if (hlinkp)
1111                 *hlinkp = 0;
1112         parent = hammer2_inode_lock_ex(dip);
1113         chain = hammer2_chain_lookup(&parent, &key_next,
1114                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1115                                      &cache_index, 0);
1116         while (chain) {
1117                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1118                     name_len == chain->data->ipdata.name_len &&
1119                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1120                         break;
1121                 }
1122                 chain = hammer2_chain_next(&parent, chain, &key_next,
1123                                            key_next,
1124                                            lhc + HAMMER2_DIRHASH_LOMASK,
1125                                            &cache_index, 0);
1126         }
1127         hammer2_inode_unlock_ex(dip, NULL);     /* retain parent */
1128
1129         /*
1130          * Not found or wrong type (isdir < 0 disables the type check).
1131          * If a hardlink pointer, type checks use the hardlink target.
1132          */
1133         if (chain == NULL) {
1134                 error = ENOENT;
1135                 goto done;
1136         }
1137         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1138                 if (hlinkp)
1139                         *hlinkp = 1;
1140                 type = chain->data->ipdata.target_type;
1141         }
1142
1143         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1144                 error = ENOTDIR;
1145                 goto done;
1146         }
1147         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1148                 error = EISDIR;
1149                 goto done;
1150         }
1151
1152         /*
1153          * Hardlink must be resolved.  We can't hold parent locked while we
1154          * do this or we could deadlock.
1155          *
1156          * On success chain will be adjusted to point at the hardlink target
1157          * and ochain will point to the hardlink pointer in the original
1158          * directory.  Otherwise chain remains pointing to the original.
1159          */
1160         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1161                 hammer2_chain_unlock(parent);
1162                 parent = NULL;
1163                 error = hammer2_hardlink_find(dip, &chain, &ochain);
1164         }
1165
1166         /*
1167          * If this is a directory the directory must be empty.  However, if
1168          * isdir < 0 we are doing a rename and the directory does not have
1169          * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1170          * and the directory does not have to be empty.
1171          *
1172          * NOTE: We check the full key range here which covers both visible
1173          *       and invisible entries.  Theoretically there should be no
1174          *       invisible (hardlink target) entries if there are no visible
1175          *       entries.
1176          */
1177         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1178                 dparent = hammer2_chain_lookup_init(chain, 0);
1179                 dchain = hammer2_chain_lookup(&dparent, &key_dummy,
1180                                               0, (hammer2_key_t)-1,
1181                                               &cache_index,
1182                                               HAMMER2_LOOKUP_NODATA);
1183                 if (dchain) {
1184                         hammer2_chain_unlock(dchain);
1185                         hammer2_chain_lookup_done(dparent);
1186                         error = ENOTEMPTY;
1187                         goto done;
1188                 }
1189                 hammer2_chain_lookup_done(dparent);
1190                 dparent = NULL;
1191                 /* dchain NULL */
1192         }
1193
1194         /*
1195          * Ok, we can now unlink the chain.  We always decrement nlinks even
1196          * if the entry can be deleted in case someone has the file open and
1197          * does an fstat().
1198          *
1199          * The chain itself will no longer be in the on-media topology but
1200          * can still be flushed to the media (e.g. if an open descriptor
1201          * remains).  When the last vnode/ip ref goes away the chain will
1202          * be marked unmodified, avoiding any further (now unnecesary) I/O.
1203          *
1204          * A non-NULL ochain indicates a hardlink.
1205          */
1206         if (ochain) {
1207                 /*
1208                  * Delete the original hardlink pointer.
1209                  *
1210                  * NOTE: parent from above is NULL when ochain != NULL
1211                  *       so we can reuse it.
1212                  */
1213                 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1214                 hammer2_chain_delete(trans, ochain, 0);
1215                 hammer2_chain_unlock(ochain);
1216
1217                 /*
1218                  * Then decrement nlinks on hardlink target, deleting
1219                  * the target when nlinks drops to 0.
1220                  */
1221                 hammer2_chain_modify(trans, &chain, 0);
1222                 --chain->data->ipdata.nlinks;
1223                 if (chain->data->ipdata.nlinks == 0)
1224                         hammer2_chain_delete(trans, chain, 0);
1225         } else {
1226                 /*
1227                  * Otherwise this was not a hardlink and we can just
1228                  * remove the entry and decrement nlinks.
1229                  *
1230                  * NOTE: *_get() integrates chain's lock into the inode lock.
1231                  */
1232                 hammer2_chain_modify(trans, &chain, 0);
1233                 ipdata = &chain->data->ipdata;
1234                 --ipdata->nlinks;
1235                 hammer2_chain_delete(trans, chain, 0);
1236         }
1237
1238         error = 0;
1239 done:
1240         if (chain)
1241                 hammer2_chain_unlock(chain);
1242         if (parent)
1243                 hammer2_chain_lookup_done(parent);
1244         if (ochain)
1245                 hammer2_chain_drop(ochain);
1246
1247         return error;
1248 }
1249
1250 /*
1251  * Given an exclusively locked inode we consolidate its chain for hardlink
1252  * creation, adding (nlinks) to the file's link count and potentially
1253  * relocating the inode to a directory common to ip->pip and tdip.
1254  *
1255  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1256  * and returning a new locked chain.
1257  *
1258  * NOTE!  This function will also replace ip->chain.
1259  */
1260 int
1261 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1262                              hammer2_chain_t **chainp,
1263                              hammer2_inode_t *tdip, int nlinks)
1264 {
1265         hammer2_inode_data_t *ipdata;
1266         hammer2_inode_t *fdip;
1267         hammer2_inode_t *cdip;
1268         hammer2_chain_t *chain;
1269         hammer2_chain_t *nchain;
1270         int error;
1271
1272         chain = *chainp;
1273         if (nlinks == 0 &&                      /* no hardlink needed */
1274             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1275                 return (0);
1276         }
1277         if (hammer2_hardlink_enable < 0) {      /* fake hardlinks */
1278                 return (0);
1279         }
1280
1281         if (hammer2_hardlink_enable == 0) {     /* disallow hardlinks */
1282                 hammer2_chain_unlock(chain);
1283                 *chainp = NULL;
1284                 return (ENOTSUP);
1285         }
1286
1287         /*
1288          * cdip will be returned with a ref, but not locked.
1289          */
1290         fdip = ip->pip;
1291         cdip = hammer2_inode_common_parent(fdip, tdip);
1292
1293         /*
1294          * If no change in the hardlink's target directory is required and
1295          * this is already a hardlink target, all we need to do is adjust
1296          * the link count.
1297          *
1298          * XXX The common parent is a big wiggly due to duplication from
1299          *     renames.  Compare the core (RBTREE) pointer instead of the
1300          *     ip's.
1301          */
1302         if (cdip == fdip &&
1303             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1304                 if (nlinks) {
1305                         hammer2_chain_modify(trans, &chain, 0);
1306                         chain->data->ipdata.nlinks += nlinks;
1307                 }
1308                 error = 0;
1309                 goto done;
1310         }
1311
1312         /*
1313          * We either have to move an existing hardlink target or we have
1314          * to create a fresh hardlink target.
1315          *
1316          * Hardlink targets are hidden inodes in a parent directory common
1317          * to all directory entries referencing the hardlink.
1318          */
1319         nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1320
1321         if (error == 0) {
1322                 /*
1323                  * Bump nlinks on duplicated hidden inode, repoint
1324                  * ip->chain.
1325                  */
1326                 hammer2_chain_modify(trans, &nchain, 0);
1327                 nchain->data->ipdata.nlinks += nlinks;
1328                 hammer2_inode_repoint(ip, cdip, nchain);
1329
1330                 /*
1331                  * If the old chain is not a hardlink target then replace
1332                  * it with a OBJTYPE_HARDLINK pointer.
1333                  *
1334                  * If the old chain IS a hardlink target then delete it.
1335                  */
1336                 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1337                         /*
1338                          * Replace original non-hardlink that's been dup'd
1339                          * with a special hardlink directory entry.  We must
1340                          * set the DIRECTDATA flag to prevent sub-chains
1341                          * from trying to synchronize to the inode if the
1342                          * file is extended afterwords.
1343                          *
1344                          * DELDUP_RECORE causes the new chain to NOT inherit
1345                          * the old chain's core (sub-tree).
1346                          */
1347                         /*hammer2_chain_modify(trans, &chain, 0);*/
1348                         hammer2_chain_delete_duplicate(trans, &chain,
1349                                                        HAMMER2_DELDUP_RECORE);
1350                         hammer2_chain_modify(trans, &chain, 0);
1351                         ipdata = &chain->data->ipdata;
1352                         ipdata->target_type = ipdata->type;
1353                         ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1354                         ipdata->uflags = 0;
1355                         ipdata->rmajor = 0;
1356                         ipdata->rminor = 0;
1357                         ipdata->ctime = 0;
1358                         ipdata->mtime = 0;
1359                         ipdata->atime = 0;
1360                         ipdata->btime = 0;
1361                         bzero(&ipdata->uid, sizeof(ipdata->uid));
1362                         bzero(&ipdata->gid, sizeof(ipdata->gid));
1363                         ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1364                         ipdata->cap_flags = 0;
1365                         ipdata->mode = 0;
1366                         ipdata->size = 0;
1367                         ipdata->nlinks = 1;
1368                         ipdata->iparent = 0;    /* XXX */
1369                         ipdata->pfs_type = 0;
1370                         ipdata->pfs_inum = 0;
1371                         bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1372                         bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1373                         ipdata->data_quota = 0;
1374                         ipdata->data_count = 0;
1375                         ipdata->inode_quota = 0;
1376                         ipdata->inode_count = 0;
1377                         ipdata->attr_tid = 0;
1378                         ipdata->dirent_tid = 0;
1379                         bzero(&ipdata->u, sizeof(ipdata->u));
1380                         /* XXX transaction ids */
1381                 } else {
1382                         hammer2_chain_delete(trans, chain, 0);
1383                 }
1384
1385                 /*
1386                  * Return the new chain.
1387                  */
1388                 hammer2_chain_unlock(chain);
1389                 chain = nchain;
1390         } else {
1391                 /*
1392                  * Return an error
1393                  */
1394                 hammer2_chain_unlock(chain);
1395                 chain = NULL;
1396         }
1397
1398         /*
1399          * Cleanup, chain/nchain already dealt with.
1400          */
1401 done:
1402         *chainp = chain;
1403         hammer2_inode_drop(cdip);
1404
1405         return (error);
1406 }
1407
1408 /*
1409  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1410  * inode while (*chainp) points to the resolved (hidden hardlink
1411  * target) inode.  In this situation when nlinks is 1 we wish to
1412  * deconsolidate the hardlink, moving it back to the directory that now
1413  * represents the only remaining link.
1414  */
1415 int
1416 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1417                                hammer2_inode_t *dip,
1418                                hammer2_chain_t **chainp,
1419                                hammer2_chain_t **ochainp)
1420 {
1421         if (*ochainp == NULL)
1422                 return (0);
1423         /* XXX */
1424         return (0);
1425 }
1426
1427 /*
1428  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1429  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1430  * the *chainp and return a new locked *chainp representing the file target
1431  * (the original *chainp will be unlocked).
1432  *
1433  * When a match is found the chain representing the original HARDLINK
1434  * will be returned in *ochainp with a ref, but not locked.
1435  *
1436  * When no match is found *chainp is set to NULL and EIO is returned.
1437  * (*ochainp) will still be set to the original chain with a ref but not
1438  * locked.
1439  */
1440 int
1441 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1442                       hammer2_chain_t **ochainp)
1443 {
1444         hammer2_chain_t *chain = *chainp;
1445         hammer2_chain_t *parent;
1446         hammer2_inode_t *ip;
1447         hammer2_inode_t *pip;
1448         hammer2_key_t key_dummy;
1449         hammer2_key_t lhc;
1450         int cache_index = -1;
1451
1452         pip = dip;
1453         hammer2_inode_ref(pip);         /* for loop */
1454         hammer2_chain_ref(chain);       /* for (*ochainp) */
1455         *ochainp = chain;
1456
1457         /*
1458          * Locate the hardlink.  pip is referenced and not locked,
1459          * ipp.
1460          *
1461          * chain is reused.
1462          */
1463         lhc = chain->data->ipdata.inum;
1464         hammer2_chain_unlock(chain);
1465         chain = NULL;
1466
1467         while ((ip = pip) != NULL) {
1468                 parent = hammer2_inode_lock_ex(ip);
1469                 hammer2_inode_drop(ip);                 /* loop */
1470                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1471                 chain = hammer2_chain_lookup(&parent, &key_dummy,
1472                                              lhc, lhc, &cache_index, 0);
1473                 hammer2_chain_lookup_done(parent);      /* discard parent */
1474                 if (chain)
1475                         break;
1476                 pip = ip->pip;          /* safe, ip held locked */
1477                 if (pip)
1478                         hammer2_inode_ref(pip);         /* loop */
1479                 hammer2_inode_unlock_ex(ip, NULL);
1480         }
1481
1482         /*
1483          * chain is locked, ip is locked.  Unlock ip, return the locked
1484          * chain.  *ipp is already set w/a ref count and not locked.
1485          *
1486          * (parent is already unlocked).
1487          */
1488         if (ip)
1489                 hammer2_inode_unlock_ex(ip, NULL);
1490         *chainp = chain;
1491         if (chain) {
1492                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1493                 /* already locked */
1494                 return (0);
1495         } else {
1496                 return (EIO);
1497         }
1498 }
1499
1500 /*
1501  * Find the directory common to both fdip and tdip, hold and return
1502  * its inode.
1503  */
1504 hammer2_inode_t *
1505 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1506 {
1507         hammer2_inode_t *scan1;
1508         hammer2_inode_t *scan2;
1509
1510         /*
1511          * We used to have a depth field but it complicated matters too
1512          * much for directory renames.  So now its ugly.  Check for
1513          * simple cases before giving up and doing it the expensive way.
1514          *
1515          * XXX need a bottom-up topology stability lock
1516          */
1517         if (fdip == tdip || fdip == tdip->pip) {
1518                 hammer2_inode_ref(fdip);
1519                 return(fdip);
1520         }
1521         if (fdip->pip == tdip) {
1522                 hammer2_inode_ref(tdip);
1523                 return(tdip);
1524         }
1525
1526         /*
1527          * XXX not MPSAFE
1528          */
1529         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1530                 scan2 = tdip;
1531                 while (scan2->pmp == tdip->pmp) {
1532                         if (scan1 == scan2) {
1533                                 hammer2_inode_ref(scan1);
1534                                 return(scan1);
1535                         }
1536                         scan2 = scan2->pip;
1537                         if (scan2 == NULL)
1538                                 break;
1539                 }
1540         }
1541         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1542               fdip, tdip);
1543         /* NOT REACHED */
1544         return(NULL);
1545 }
1546
1547 /*
1548  * Synchronize the inode's frontend state with the chain state prior
1549  * to any explicit flush of the inode or any strategy write call.
1550  *
1551  * Called with a locked inode.
1552  */
1553 void
1554 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip, 
1555                     hammer2_chain_t **chainp)
1556 {
1557         hammer2_inode_data_t *ipdata;
1558         hammer2_chain_t *parent;
1559         hammer2_chain_t *chain;
1560         hammer2_key_t lbase;
1561         hammer2_key_t key_next;
1562         int cache_index;
1563
1564         ipdata = &ip->chain->data->ipdata;
1565
1566         if (ip->flags & HAMMER2_INODE_MTIME) {
1567                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1568                 atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1569                 ipdata->mtime = ip->mtime;
1570         }
1571         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1572                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1573                 ipdata->size = ip->size;
1574                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1575
1576                 /*
1577                  * We must delete any chains beyond the EOF.  The chain
1578                  * straddling the EOF will be pending in the bioq.
1579                  */
1580                 lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1581                         ~HAMMER2_PBUFMASK64;
1582                 parent = hammer2_chain_lookup_init(ip->chain, 0);
1583                 chain = hammer2_chain_lookup(&parent, &key_next,
1584                                              lbase, (hammer2_key_t)-1,
1585                                              &cache_index,
1586                                              HAMMER2_LOOKUP_NODATA);
1587                 while (chain) {
1588                         /*
1589                          * Degenerate embedded case, nothing to loop on
1590                          */
1591                         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1592                                 hammer2_chain_unlock(chain);
1593                                 break;
1594                         }
1595                         if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1596                                 hammer2_chain_delete(trans, chain, 0);
1597                         }
1598                         chain = hammer2_chain_next(&parent, chain, &key_next,
1599                                                    key_next, (hammer2_key_t)-1,
1600                                                    &cache_index,
1601                                                    HAMMER2_LOOKUP_NODATA);
1602                 }
1603                 hammer2_chain_lookup_done(parent);
1604         } else
1605         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1606                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1607                 ipdata->size = ip->size;
1608                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1609
1610                 /*
1611                  * When resizing larger we may not have any direct-data
1612                  * available.
1613                  */
1614                 if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1615                     ip->size > HAMMER2_EMBEDDED_BYTES) {
1616                         ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1617                         bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
1618                 }
1619         }
1620 }