a4eff73b37023c815fa5a4e40afa1c4054337e9c
[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  * The underlying chain is also locked and returned.
67  *
68  * NOTE: We don't combine the inode/chain lock because putting away an
69  *       inode would otherwise confuse multiple lock holders of the inode.
70  */
71 hammer2_chain_t *
72 hammer2_inode_lock_ex(hammer2_inode_t *ip)
73 {
74         hammer2_chain_t *chain;
75
76         hammer2_inode_ref(ip);
77         ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
78
79         /*
80          * ip->chain fixup.  Certain duplications used to move inodes
81          * into indirect blocks (for example) can cause ip->chain to
82          * become stale.
83          */
84 again:
85         chain = ip->chain;
86         if (hammer2_chain_refactor_test(chain, 1)) {
87                 spin_lock(&chain->core->cst.spin);
88                 while (hammer2_chain_refactor_test(chain, 1))
89                         chain = chain->next_parent;
90                 if (ip->chain != chain) {
91                         hammer2_chain_ref(chain);
92                         spin_unlock(&chain->core->cst.spin);
93                         hammer2_inode_repoint(ip, NULL, chain);
94                         hammer2_chain_drop(chain);
95                 } else {
96                         spin_unlock(&chain->core->cst.spin);
97                 }
98         }
99
100         KKASSERT(chain != NULL);        /* for now */
101         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
102
103         /*
104          * Resolve duplication races
105          */
106         if (hammer2_chain_refactor_test(chain, 1)) {
107                 hammer2_chain_unlock(chain);
108                 goto again;
109         }
110         return (chain);
111 }
112
113 void
114 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
115 {
116         /*
117          * XXX this will catch parent directories too which we don't
118          *     really want.
119          */
120         if (chain)
121                 hammer2_chain_unlock(chain);
122         ccms_thread_unlock(&ip->topo_cst);
123         hammer2_inode_drop(ip);
124 }
125
126 /*
127  * NOTE: We don't combine the inode/chain lock because putting away an
128  *       inode would otherwise confuse multiple lock holders of the inode.
129  *
130  *       Shared locks are especially sensitive to having too many shared
131  *       lock counts (from the same thread) on certain paths which might
132  *       need to upgrade them.  Only one count of a shared lock can be
133  *       upgraded.
134  */
135 hammer2_chain_t *
136 hammer2_inode_lock_sh(hammer2_inode_t *ip)
137 {
138         hammer2_chain_t *chain;
139
140         hammer2_inode_ref(ip);
141 again:
142         ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
143
144         chain = ip->chain;
145         KKASSERT(chain != NULL);        /* for now */
146         hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
147                                   HAMMER2_RESOLVE_SHARED);
148
149         /*
150          * Resolve duplication races
151          */
152         if (hammer2_chain_refactor_test(chain, 1)) {
153                 hammer2_chain_unlock(chain);
154                 ccms_thread_unlock(&ip->topo_cst);
155                 chain = hammer2_inode_lock_ex(ip);
156                 hammer2_inode_unlock_ex(ip, chain);
157                 goto again;
158         }
159         return (chain);
160 }
161
162 void
163 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
164 {
165         if (chain)
166                 hammer2_chain_unlock(chain);
167         ccms_thread_unlock(&ip->topo_cst);
168         hammer2_inode_drop(ip);
169 }
170
171 ccms_state_t
172 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
173 {
174         return(ccms_thread_lock_temp_release(&ip->topo_cst));
175 }
176
177 void
178 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
179 {
180         ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
181 }
182
183 ccms_state_t
184 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
185 {
186         return(ccms_thread_lock_upgrade(&ip->topo_cst));
187 }
188
189 void
190 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
191 {
192         ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
193 }
194
195 /*
196  * Lookup an inode by inode number
197  */
198 hammer2_inode_t *
199 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
200 {
201         hammer2_inode_t *ip;
202
203         if (pmp) {
204                 spin_lock(&pmp->inum_spin);
205                 ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
206                 if (ip)
207                         hammer2_inode_ref(ip);
208                 spin_unlock(&pmp->inum_spin);
209         } else {
210                 ip = NULL;
211         }
212         return(ip);
213 }
214
215 /*
216  * Adding a ref to an inode is only legal if the inode already has at least
217  * one ref.
218  */
219 void
220 hammer2_inode_ref(hammer2_inode_t *ip)
221 {
222         atomic_add_int(&ip->refs, 1);
223 }
224
225 /*
226  * Drop an inode reference, freeing the inode when the last reference goes
227  * away.
228  */
229 void
230 hammer2_inode_drop(hammer2_inode_t *ip)
231 {
232         hammer2_mount_t *hmp;
233         hammer2_inode_t *pip;
234         u_int refs;
235
236         while (ip) {
237                 refs = ip->refs;
238                 cpu_ccfence();
239                 if (refs == 1) {
240                         /*
241                          * Transition to zero, must interlock with
242                          * the inode inumber lookup tree (if applicable).
243                          *
244                          * NOTE: The super-root inode has no pmp.
245                          */
246                         if (ip->pmp)
247                                 spin_lock(&ip->pmp->inum_spin);
248
249                         if (atomic_cmpset_int(&ip->refs, 1, 0)) {
250                                 KKASSERT(ip->topo_cst.count == 0);
251                                 if (ip->flags & HAMMER2_INODE_ONRBTREE) {
252                                         atomic_clear_int(&ip->flags,
253                                                      HAMMER2_INODE_ONRBTREE);
254                                         RB_REMOVE(hammer2_inode_tree,
255                                                   &ip->pmp->inum_tree,
256                                                   ip);
257                                 }
258                                 if (ip->pmp)
259                                         spin_unlock(&ip->pmp->inum_spin);
260
261                                 hmp = ip->hmp;
262                                 ip->hmp = NULL;
263                                 pip = ip->pip;
264                                 ip->pip = 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                                 kfree(ip, hmp->minode);
278                                 ip = pip;
279                                 /* continue with pip (can be NULL) */
280                         } else {
281                                 if (ip->pmp)
282                                         spin_unlock(&ip->pmp->inum_spin);
283                         }
284                 } else {
285                         /*
286                          * Non zero transition
287                          */
288                         if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
289                                 break;
290                 }
291         }
292 }
293
294 /*
295  * Get the vnode associated with the given inode, allocating the vnode if
296  * necessary.  The vnode will be returned exclusively locked.
297  *
298  * The caller must lock the inode (shared or exclusive).
299  *
300  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
301  * races.
302  */
303 struct vnode *
304 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
305 {
306         hammer2_inode_data_t *ipdata;
307         hammer2_pfsmount_t *pmp;
308         struct vnode *vp;
309         ccms_state_t ostate;
310
311         pmp = ip->pmp;
312         KKASSERT(pmp != NULL);
313         *errorp = 0;
314         ipdata = &ip->chain->data->ipdata;
315
316         for (;;) {
317                 /*
318                  * Attempt to reuse an existing vnode assignment.  It is
319                  * possible to race a reclaim so the vget() may fail.  The
320                  * inode must be unlocked during the vget() to avoid a
321                  * deadlock against a reclaim.
322                  */
323                 vp = ip->vp;
324                 if (vp) {
325                         /*
326                          * Inode must be unlocked during the vget() to avoid
327                          * possible deadlocks, but leave the ip ref intact.
328                          *
329                          * vnode is held to prevent destruction during the
330                          * vget().  The vget() can still fail if we lost
331                          * a reclaim race on the vnode.
332                          */
333                         vhold_interlocked(vp);
334                         ostate = hammer2_inode_lock_temp_release(ip);
335                         if (vget(vp, LK_EXCLUSIVE)) {
336                                 vdrop(vp);
337                                 hammer2_inode_lock_temp_restore(ip, ostate);
338                                 continue;
339                         }
340                         hammer2_inode_lock_temp_restore(ip, ostate);
341                         vdrop(vp);
342                         /* vp still locked and ref from vget */
343                         if (ip->vp != vp) {
344                                 kprintf("hammer2: igetv race %p/%p\n",
345                                         ip->vp, vp);
346                                 vput(vp);
347                                 continue;
348                         }
349                         *errorp = 0;
350                         break;
351                 }
352
353                 /*
354                  * No vnode exists, allocate a new vnode.  Beware of
355                  * allocation races.  This function will return an
356                  * exclusively locked and referenced vnode.
357                  */
358                 *errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
359                 if (*errorp) {
360                         kprintf("hammer2: igetv getnewvnode failed %d\n",
361                                 *errorp);
362                         vp = NULL;
363                         break;
364                 }
365
366                 /*
367                  * Lock the inode and check for an allocation race.
368                  */
369                 ostate = hammer2_inode_lock_upgrade(ip);
370                 if (ip->vp != NULL) {
371                         vp->v_type = VBAD;
372                         vx_put(vp);
373                         hammer2_inode_lock_downgrade(ip, ostate);
374                         continue;
375                 }
376
377                 switch (ipdata->type) {
378                 case HAMMER2_OBJTYPE_DIRECTORY:
379                         vp->v_type = VDIR;
380                         break;
381                 case HAMMER2_OBJTYPE_REGFILE:
382                         vp->v_type = VREG;
383                         vinitvmio(vp, ipdata->size,
384                                   HAMMER2_LBUFSIZE,
385                                   (int)ipdata->size & HAMMER2_LBUFMASK);
386                         break;
387                 case HAMMER2_OBJTYPE_SOFTLINK:
388                         /*
389                          * XXX for now we are using the generic file_read
390                          * and file_write code so we need a buffer cache
391                          * association.
392                          */
393                         vp->v_type = VLNK;
394                         vinitvmio(vp, ipdata->size,
395                                   HAMMER2_LBUFSIZE,
396                                   (int)ipdata->size & HAMMER2_LBUFMASK);
397                         break;
398                 /* XXX FIFO */
399                 default:
400                         panic("hammer2: unhandled objtype %d", ipdata->type);
401                         break;
402                 }
403
404                 if (ip == pmp->iroot)
405                         vsetflags(vp, VROOT);
406
407                 vp->v_data = ip;
408                 ip->vp = vp;
409                 hammer2_inode_ref(ip);          /* vp association */
410                 hammer2_inode_lock_downgrade(ip, ostate);
411                 break;
412         }
413
414         /*
415          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
416          */
417         if (hammer2_debug & 0x0002) {
418                 kprintf("igetv vp %p refs %d aux %d\n",
419                         vp, vp->v_sysref.refcnt, vp->v_auxrefs);
420         }
421         return (vp);
422 }
423
424 /*
425  * The passed-in chain must be locked and the returned inode will also be
426  * locked.  This routine typically locates or allocates the inode, assigns
427  * ip->chain (adding a ref to chain if necessary), and returns the inode.
428  *
429  * The hammer2_inode structure regulates the interface between the high level
430  * kernel VNOPS API and the filesystem backend (the chains).
431  *
432  * WARNING!  This routine sucks up the chain's lock (makes it part of the
433  *           inode lock from the point of view of the inode lock API),
434  *           so callers need to be careful.
435  *
436  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
437  *           is allowed to pass pmp == NULL and dip == NULL for sroot.
438  */
439 hammer2_inode_t *
440 hammer2_inode_get(hammer2_mount_t *hmp, hammer2_pfsmount_t *pmp,
441                   hammer2_inode_t *dip, hammer2_chain_t *chain)
442 {
443         hammer2_inode_t *nip;
444
445         KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
446
447         /*
448          * Interlocked lookup/ref of the inode.  This code is only needed
449          * when looking up inodes with nlinks != 0 (TODO: optimize out
450          * otherwise and test for duplicates).
451          */
452 again:
453         for (;;) {
454                 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
455                 if (nip == NULL)
456                         break;
457                 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
458                 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
459                         ccms_thread_unlock(&nip->topo_cst);
460                         hammer2_inode_drop(nip);
461                         continue;
462                 }
463                 if (nip->chain != chain)
464                         hammer2_inode_repoint(nip, NULL, chain);
465
466                 /*
467                  * Consolidated nip/nip->chain is locked (chain locked
468                  * by caller).
469                  */
470                 return nip;
471         }
472
473         /*
474          * We couldn't find the inode number, create a new inode.
475          */
476         nip = kmalloc(sizeof(*nip), hmp->minode, M_WAITOK | M_ZERO);
477         nip->inum = chain->data->ipdata.inum;
478         hammer2_inode_repoint(nip, NULL, chain);
479         nip->pip = dip;                         /* can be NULL */
480         if (dip)
481                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
482
483         nip->pmp = pmp;
484         nip->hmp = hmp;
485
486         /*
487          * ref and lock on nip gives it state compatible to after a
488          * hammer2_inode_lock_ex() call.
489          */
490         nip->refs = 1;
491         ccms_cst_init(&nip->topo_cst, &nip->chain);
492         ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
493         /* combination of thread lock and chain lock == inode lock */
494
495         /*
496          * Attempt to add the inode.  If it fails we raced another inode
497          * get.  Undo all the work and try again.
498          */
499         if (pmp) {
500                 spin_lock(&pmp->inum_spin);
501                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
502                         spin_unlock(&pmp->inum_spin);
503                         ccms_thread_unlock(&nip->topo_cst);
504                         hammer2_inode_drop(nip);
505                         goto again;
506                 }
507                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
508                 spin_unlock(&pmp->inum_spin);
509         }
510
511         return (nip);
512 }
513
514 /*
515  * Create a new inode in the specified directory using the vattr to
516  * figure out the type of inode.
517  *
518  * If no error occurs the new inode with its chain locked is returned in
519  * *nipp, otherwise an error is returned and *nipp is set to NULL.
520  *
521  * If vap and/or cred are NULL the related fields are not set and the
522  * inode type defaults to a directory.  This is used when creating PFSs
523  * under the super-root, so the inode number is set to 1 in this case.
524  *
525  * dip is not locked on entry.
526  */
527 hammer2_inode_t *
528 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
529                      struct vattr *vap, struct ucred *cred,
530                      const uint8_t *name, size_t name_len,
531                      hammer2_chain_t **chainp, int *errorp)
532 {
533         hammer2_inode_data_t *dipdata;
534         hammer2_inode_data_t *nipdata;
535         hammer2_mount_t *hmp;
536         hammer2_chain_t *chain;
537         hammer2_chain_t *parent;
538         hammer2_inode_t *nip;
539         hammer2_key_t lhc;
540         int error;
541         uid_t xuid;
542         uuid_t dip_uid;
543         uuid_t dip_gid;
544         uint32_t dip_mode;
545
546         hmp = dip->hmp;
547         lhc = hammer2_dirhash(name, name_len);
548         *errorp = 0;
549
550         /*
551          * Locate the inode or indirect block to create the new
552          * entry in.  At the same time check for key collisions
553          * and iterate until we don't get one.
554          *
555          * NOTE: hidden inodes do not have iterators.
556          */
557 retry:
558         parent = hammer2_inode_lock_ex(dip);
559         dipdata = &dip->chain->data->ipdata;
560         dip_uid = dipdata->uid;
561         dip_gid = dipdata->gid;
562         dip_mode = dipdata->mode;
563
564         error = 0;
565         while (error == 0) {
566                 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
567                 if (chain == NULL)
568                         break;
569                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
570                         error = ENOSPC;
571                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
572                         error = ENOSPC;
573                 hammer2_chain_unlock(chain);
574                 chain = NULL;
575                 ++lhc;
576         }
577         if (error == 0) {
578                 error = hammer2_chain_create(trans, &parent, &chain,
579                                              lhc, 0,
580                                              HAMMER2_BREF_TYPE_INODE,
581                                              HAMMER2_INODE_BYTES);
582         }
583
584         /*
585          * Cleanup and handle retries.
586          */
587         if (error == EAGAIN) {
588                 hammer2_chain_ref(parent);
589                 hammer2_inode_unlock_ex(dip, parent);
590                 hammer2_chain_wait(parent);
591                 hammer2_chain_drop(parent);
592                 goto retry;
593         }
594         hammer2_inode_unlock_ex(dip, parent);
595
596         if (error) {
597                 KKASSERT(chain == NULL);
598                 *errorp = error;
599                 return (NULL);
600         }
601
602         /*
603          * Set up the new inode.
604          *
605          * NOTE: *_get() integrates chain's lock into the inode lock.
606          *
607          * NOTE: Only one new inode can currently be created per
608          *       transaction.  If the need arises we can adjust
609          *       hammer2_trans_init() to allow more.
610          */
611         chain->data->ipdata.inum = trans->sync_tid;
612         nip = hammer2_inode_get(dip->hmp, dip->pmp, dip, chain);
613         nipdata = &chain->data->ipdata;
614
615         if (vap) {
616                 KKASSERT(trans->inodes_created == 0);
617                 nipdata->type = hammer2_get_obj_type(vap->va_type);
618                 nipdata->inum = trans->sync_tid;
619                 ++trans->inodes_created;
620         } else {
621                 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
622                 nipdata->inum = 1;
623         }
624         nipdata->version = HAMMER2_INODE_VERSION_ONE;
625         hammer2_update_time(&nipdata->ctime);
626         nipdata->mtime = nipdata->ctime;
627         if (vap)
628                 nipdata->mode = vap->va_mode;
629         nipdata->nlinks = 1;
630         if (vap) {
631                 if (dip) {
632                         xuid = hammer2_to_unix_xid(&dip_uid);
633                         xuid = vop_helper_create_uid(dip->pmp->mp,
634                                                      dip_mode,
635                                                      xuid,
636                                                      cred,
637                                                      &vap->va_mode);
638                 } else {
639                         xuid = 0;
640                 }
641                 if (vap->va_vaflags & VA_UID_UUID_VALID)
642                         nipdata->uid = vap->va_uid_uuid;
643                 else if (vap->va_uid != (uid_t)VNOVAL)
644                         hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
645                 else
646                         hammer2_guid_to_uuid(&nipdata->uid, xuid);
647
648                 if (vap->va_vaflags & VA_GID_UUID_VALID)
649                         nipdata->gid = vap->va_gid_uuid;
650                 else if (vap->va_gid != (gid_t)VNOVAL)
651                         hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
652                 else if (dip)
653                         nipdata->gid = dip_gid;
654         }
655
656         /*
657          * Regular files and softlinks allow a small amount of data to be
658          * directly embedded in the inode.  This flag will be cleared if
659          * the size is extended past the embedded limit.
660          */
661         if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
662             nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
663                 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
664         }
665
666         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
667         bcopy(name, nipdata->filename, name_len);
668         nipdata->name_key = lhc;
669         nipdata->name_len = name_len;
670         *chainp = chain;
671
672         return (nip);
673 }
674
675 /*
676  * chain may have been moved around by the create.
677  */
678 static
679 void
680 hammer2_chain_refactor(hammer2_chain_t **chainp)
681 {
682         hammer2_chain_t *chain = *chainp;
683         hammer2_chain_core_t *core;
684
685         core = chain->core;
686         spin_lock(&core->cst.spin);
687         while (hammer2_chain_refactor_test(chain, 1)) {
688                 chain = chain->next_parent;
689                 while (hammer2_chain_refactor_test(chain, 1))
690                         chain = chain->next_parent;
691                 hammer2_chain_ref(chain);
692                 spin_unlock(&core->cst.spin);
693
694                 hammer2_chain_unlock(*chainp);
695                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
696                                           HAMMER2_RESOLVE_NOREF); /* eat ref */
697                 *chainp = chain;
698                 spin_lock(&core->cst.spin);
699         }
700         spin_unlock(&core->cst.spin);
701 }
702
703 /*
704  * ochain represents the target file inode.  We need to move it to the
705  * specified common parent directory (dip) and rename it to a special
706  * invisible "0xINODENUMBER" filename.
707  *
708  * We use chain_duplicate and duplicate ochain at the new location,
709  * renaming it appropriately.  We create a temporary chain and
710  * then delete it to placemark where the duplicate will go.  Both of
711  * these use the inode number for (lhc) (the key), generating the
712  * invisible filename.
713  */
714 static
715 hammer2_chain_t *
716 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
717                         hammer2_inode_t *dip, int *errorp)
718 {
719         hammer2_inode_data_t *nipdata;
720         hammer2_mount_t *hmp;
721         hammer2_chain_t *parent;
722         hammer2_chain_t *ochain;
723         hammer2_chain_t *nchain;
724         hammer2_chain_t *tmp;
725         hammer2_key_t lhc;
726         hammer2_blockref_t bref;
727
728         ochain = *ochainp;
729         *errorp = 0;
730         hmp = dip->hmp;
731         lhc = ochain->data->ipdata.inum;
732         KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
733
734         /*
735          * Locate the inode or indirect block to create the new
736          * entry in.  lhc represents the inode number so there is
737          * no collision iteration.
738          *
739          * There should be no key collisions with invisible inode keys.
740          */
741 retry:
742         parent = hammer2_chain_lookup_init(dip->chain, 0);
743         nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
744         if (nchain) {
745                 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
746                         nchain, parent, dip, dip->chain);
747                 hammer2_chain_unlock(nchain);
748                 nchain = NULL;
749                 *errorp = ENOSPC;
750 #if 1
751                 Debugger("X3");
752 #endif
753         }
754
755         /*
756          * Create entry in common parent directory using the seek position
757          * calculated above.
758          */
759         if (*errorp == 0) {
760                 KKASSERT(nchain == NULL);
761                 *errorp = hammer2_chain_create(trans, &parent, &nchain,
762                                                lhc, 0,
763                                                HAMMER2_BREF_TYPE_INODE,/* n/a */
764                                                HAMMER2_INODE_BYTES);   /* n/a */
765                 hammer2_chain_refactor(&ochain);
766                 *ochainp = ochain;
767         }
768
769         /*
770          * Cleanup and handle retries.
771          */
772         if (*errorp == EAGAIN) {
773                 hammer2_chain_ref(parent);
774                 hammer2_chain_lookup_done(parent);
775                 hammer2_chain_wait(parent);
776                 hammer2_chain_drop(parent);
777                 goto retry;
778         }
779
780         /*
781          * Handle the error case
782          */
783         if (*errorp) {
784                 KKASSERT(nchain == NULL);
785                 hammer2_chain_lookup_done(parent);
786                 return (NULL);
787         }
788
789         /*
790          * Use chain as a placeholder for (lhc), delete it and replace
791          * it with our duplication.
792          *
793          * Gain a second lock on ochain for the duplication function to
794          * unlock, maintain the caller's original lock across the call.
795          *
796          * This is a bit messy.
797          */
798         hammer2_chain_delete(trans, nchain);
799         hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
800         tmp = ochain;
801         bref = tmp->bref;
802         bref.key = lhc;                 /* invisible dir entry key */
803         bref.keybits = 0;
804         hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
805         hammer2_chain_lookup_done(parent);
806         hammer2_chain_unlock(nchain);   /* no longer needed */
807
808         /*
809          * Now set chain to our duplicate and modify it appropriately.
810          *
811          * Directory entries are inodes but this is a hidden hardlink
812          * target.  The name isn't used but to ease debugging give it
813          * a name after its inode number.
814          */
815         nchain = tmp;
816         tmp = NULL;     /* safety */
817
818         hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
819         nipdata = &nchain->data->ipdata;
820         ksnprintf(nipdata->filename, sizeof(nipdata->filename),
821                   "0x%016jx", (intmax_t)nipdata->inum);
822         nipdata->name_len = strlen(nipdata->filename);
823         nipdata->name_key = lhc;
824
825         return (nchain);
826 }
827
828 /*
829  * Connect the target inode represented by (*chainp) to the media topology
830  * at (dip, name, len).
831  *
832  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
833  * entry instead of connecting (*chainp).
834  *
835  * If hlink is FALSE this function uses chain_duplicate() to make a copy
836  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
837  * by the caller in this case (e.g. rename).
838  */
839 int
840 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
841                       hammer2_inode_t *dip, hammer2_chain_t **chainp,
842                       const uint8_t *name, size_t name_len)
843 {
844         hammer2_inode_data_t *ipdata;
845         hammer2_mount_t *hmp;
846         hammer2_chain_t *nchain;
847         hammer2_chain_t *parent;
848         hammer2_chain_t *ochain;
849         hammer2_key_t lhc;
850         int error;
851
852         hmp = dip->hmp;
853
854         ochain = *chainp;
855
856         /*
857          * Since ochain is either disconnected from the topology or represents
858          * a hardlink terminus which is always a parent of or equal to dip,
859          * we should be able to safely lock dip->chain for our setup.
860          */
861         parent = hammer2_chain_lookup_init(dip->chain, 0);
862
863         lhc = hammer2_dirhash(name, name_len);
864
865         /*
866          * Locate the inode or indirect block to create the new
867          * entry in.  At the same time check for key collisions
868          * and iterate until we don't get one.
869          */
870         error = 0;
871         while (error == 0) {
872                 nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
873                 if (nchain == NULL)
874                         break;
875                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
876                         error = ENOSPC;
877                 hammer2_chain_unlock(nchain);
878                 nchain = NULL;
879                 ++lhc;
880         }
881
882         if (error == 0) {
883                 if (hlink) {
884                         /*
885                          * Hardlink pointer needed, create totally fresh
886                          * directory entry.
887                          */
888                         KKASSERT(nchain == NULL);
889                         error = hammer2_chain_create(trans, &parent, &nchain,
890                                                      lhc, 0,
891                                                      HAMMER2_BREF_TYPE_INODE,
892                                                      HAMMER2_INODE_BYTES);
893                         hammer2_chain_refactor(&ochain);
894                 } else {
895                         /*
896                          * Reconnect the original chain and rename.  Use
897                          * chain_duplicate().  The caller will likely delete
898                          * or has already deleted the original chain in
899                          * this case.
900                          *
901                          * NOTE: chain_duplicate() generates a new chain
902                          *       with CHAIN_DELETED cleared (ochain typically
903                          *       has it set from the file unlink).
904                          */
905                         nchain = ochain;
906                         ochain = NULL;
907                         hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
908                         error = hammer2_chain_create(trans, &parent, &nchain,
909                                                      lhc, 0,
910                                                      HAMMER2_BREF_TYPE_INODE,
911                                                      HAMMER2_INODE_BYTES);
912                 }
913         }
914
915         /*
916          * Unlock stuff.
917          */
918         KKASSERT(error != EAGAIN);
919         hammer2_chain_lookup_done(parent);
920         parent = NULL;
921
922         /*
923          * nchain should be NULL on error, leave ochain (== *chainp) alone.
924          */
925         if (error) {
926                 KKASSERT(nchain == NULL);
927                 return (error);
928         }
929
930         /*
931          * Directory entries are inodes so if the name has changed we have
932          * to update the inode.
933          *
934          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
935          * chain, the caller will access the hardlink via the actual hardlink
936          * target file and not the hardlink pointer entry, so we must still
937          * return ochain.
938          */
939         if (hlink && hammer2_hardlink_enable >= 0) {
940                 /*
941                  * Create the HARDLINK pointer.  oip represents the hardlink
942                  * target in this situation.
943                  *
944                  * We will return ochain (the hardlink target).
945                  */
946                 hammer2_chain_modify(trans, &nchain,
947                                      HAMMER2_MODIFY_ASSERTNOCOPY);
948                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
949                 ipdata = &nchain->data->ipdata;
950                 bcopy(name, ipdata->filename, name_len);
951                 ipdata->name_key = lhc;
952                 ipdata->name_len = name_len;
953                 ipdata->target_type = ochain->data->ipdata.type;
954                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
955                 ipdata->inum = ochain->data->ipdata.inum;
956                 ipdata->nlinks = 1;
957                 hammer2_chain_unlock(nchain);
958                 nchain = ochain;
959                 ochain = NULL;
960         } else if (hlink && hammer2_hardlink_enable < 0) {
961                 /*
962                  * Create a snapshot (hardlink fake mode for debugging).
963                  * (ochain already flushed above so we can just copy the
964                  * bref XXX).
965                  *
966                  * Since this is a snapshot we return nchain in the fake
967                  * hardlink case.
968                  */
969                 hammer2_chain_modify(trans, &nchain,
970                                      HAMMER2_MODIFY_ASSERTNOCOPY);
971                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
972                 ipdata = &nchain->data->ipdata;
973                 *ipdata = ochain->data->ipdata;
974                 bcopy(name, ipdata->filename, name_len);
975                 ipdata->name_key = lhc;
976                 ipdata->name_len = name_len;
977                 kprintf("created fake hardlink %*.*s\n",
978                         (int)name_len, (int)name_len, name);
979         } else {
980                 /*
981                  * nchain is a duplicate of ochain at the new location.
982                  * We must fixup the name stored in oip.  The bref key
983                  * has already been set up.
984                  */
985                 hammer2_chain_modify(trans, &nchain,
986                                      HAMMER2_MODIFY_ASSERTNOCOPY);
987                 ipdata = &nchain->data->ipdata;
988
989                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
990                 bcopy(name, ipdata->filename, name_len);
991                 ipdata->name_key = lhc;
992                 ipdata->name_len = name_len;
993                 ipdata->nlinks = 1;
994         }
995
996         /*
997          * We are replacing ochain with nchain, unlock ochain.  In the
998          * case where ochain is left unchanged the code above sets
999          * nchain to ochain and ochain to NULL, resulting in a NOP here.
1000          */
1001         if (ochain)
1002                 hammer2_chain_unlock(ochain);
1003         *chainp = nchain;
1004
1005         return (0);
1006 }
1007
1008 /*
1009  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1010  * locked.
1011  *
1012  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1013  * and nchain is ref'd.
1014  */
1015 void
1016 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1017                       hammer2_chain_t *nchain)
1018 {
1019         hammer2_chain_t *ochain;
1020         hammer2_inode_t *opip;
1021
1022         /*
1023          * Repoint ip->chain if requested.
1024          */
1025         ochain = ip->chain;
1026         ip->chain = nchain;
1027         if (nchain)
1028                 hammer2_chain_ref(nchain);
1029         if (ochain)
1030                 hammer2_chain_drop(ochain);
1031
1032         /*
1033          * Repoint ip->pip if requested (non-NULL pip).
1034          */
1035         if (pip && ip->pip != pip) {
1036                 opip = ip->pip;
1037                 hammer2_inode_ref(pip);
1038                 ip->pip = pip;
1039                 if (opip)
1040                         hammer2_inode_drop(opip);
1041         }
1042 }
1043
1044 /*
1045  * Unlink the file from the specified directory inode.  The directory inode
1046  * does not need to be locked.
1047  *
1048  * isdir determines whether a directory/non-directory check should be made.
1049  * No check is made if isdir is set to -1.
1050  *
1051  * NOTE!  This function does not prevent the underlying file from still
1052  *        being used if it has other refs (such as from an inode, or if it's
1053  *        chain is manually held).  However, the caller is responsible for
1054  *        fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
1055  */
1056 int
1057 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1058                     const uint8_t *name, size_t name_len,
1059                     int isdir, int *hlinkp)
1060 {
1061         hammer2_inode_data_t *ipdata;
1062         hammer2_mount_t *hmp;
1063         hammer2_chain_t *parent;
1064         hammer2_chain_t *ochain;
1065         hammer2_chain_t *chain;
1066         hammer2_chain_t *dparent;
1067         hammer2_chain_t *dchain;
1068         hammer2_key_t lhc;
1069         int error;
1070         uint8_t type;
1071
1072         error = 0;
1073         ochain = NULL;
1074         hmp = dip->hmp;
1075         lhc = hammer2_dirhash(name, name_len);
1076
1077         /*
1078          * Search for the filename in the directory
1079          */
1080         if (hlinkp)
1081                 *hlinkp = 0;
1082         parent = hammer2_inode_lock_ex(dip);
1083         chain = hammer2_chain_lookup(&parent,
1084                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1085                                      0);
1086         while (chain) {
1087                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1088                     name_len == chain->data->ipdata.name_len &&
1089                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1090                         break;
1091                 }
1092                 chain = hammer2_chain_next(&parent, chain,
1093                                            lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1094                                            0);
1095         }
1096         hammer2_inode_unlock_ex(dip, NULL);     /* retain parent */
1097
1098         /*
1099          * Not found or wrong type (isdir < 0 disables the type check).
1100          * If a hardlink pointer, type checks use the hardlink target.
1101          */
1102         if (chain == NULL) {
1103                 error = ENOENT;
1104                 goto done;
1105         }
1106         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1107                 if (hlinkp)
1108                         *hlinkp = 1;
1109                 type = chain->data->ipdata.target_type;
1110         }
1111
1112         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1113                 error = ENOTDIR;
1114                 goto done;
1115         }
1116         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1117                 error = EISDIR;
1118                 goto done;
1119         }
1120
1121         /*
1122          * Hardlink must be resolved.  We can't hold parent locked while we
1123          * do this or we could deadlock.
1124          *
1125          * On success chain will be adjusted to point at the hardlink target
1126          * and ochain will point to the hardlink pointer in the original
1127          * directory.  Otherwise chain remains pointing to the original.
1128          */
1129         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1130                 hammer2_chain_unlock(parent);
1131                 parent = NULL;
1132                 error = hammer2_hardlink_find(dip, &chain, &ochain);
1133         }
1134
1135         /*
1136          * If this is a directory the directory must be empty.  However, if
1137          * isdir < 0 we are doing a rename and the directory does not have
1138          * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1139          * and the directory does not have to be empty.
1140          *
1141          * NOTE: We check the full key range here which covers both visible
1142          *       and invisible entries.  Theoretically there should be no
1143          *       invisible (hardlink target) entries if there are no visible
1144          *       entries.
1145          */
1146         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1147                 dparent = hammer2_chain_lookup_init(chain, 0);
1148                 dchain = hammer2_chain_lookup(&dparent,
1149                                               0, (hammer2_key_t)-1,
1150                                               HAMMER2_LOOKUP_NODATA);
1151                 if (dchain) {
1152                         hammer2_chain_unlock(dchain);
1153                         hammer2_chain_lookup_done(dparent);
1154                         error = ENOTEMPTY;
1155                         goto done;
1156                 }
1157                 hammer2_chain_lookup_done(dparent);
1158                 dparent = NULL;
1159                 /* dchain NULL */
1160         }
1161
1162         /*
1163          * Ok, we can now unlink the chain.  We always decrement nlinks even
1164          * if the entry can be deleted in case someone has the file open and
1165          * does an fstat().
1166          *
1167          * The chain itself will no longer be in the on-media topology but
1168          * can still be flushed to the media (e.g. if an open descriptor
1169          * remains).  When the last vnode/ip ref goes away the chain will
1170          * be marked unmodified, avoiding any further (now unnecesary) I/O.
1171          *
1172          * A non-NULL ochain indicates a hardlink.
1173          */
1174         if (ochain) {
1175                 /*
1176                  * Delete the original hardlink pointer.
1177                  *
1178                  * NOTE: parent from above is NULL when ochain != NULL
1179                  *       so we can reuse it.
1180                  */
1181                 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1182                 hammer2_chain_delete(trans, ochain);
1183                 hammer2_chain_unlock(ochain);
1184
1185                 /*
1186                  * Then decrement nlinks on hardlink target, deleting
1187                  * the target when nlinks drops to 0.
1188                  */
1189                 hammer2_chain_modify(trans, &chain, 0);
1190                 --chain->data->ipdata.nlinks;
1191                 if (chain->data->ipdata.nlinks == 0)
1192                         hammer2_chain_delete(trans, chain);
1193         } else {
1194                 /*
1195                  * Otherwise this was not a hardlink and we can just
1196                  * remove the entry and decrement nlinks.
1197                  *
1198                  * NOTE: *_get() integrates chain's lock into the inode lock.
1199                  */
1200                 hammer2_chain_modify(trans, &chain, 0);
1201                 ipdata = &chain->data->ipdata;
1202                 --ipdata->nlinks;
1203                 hammer2_chain_delete(trans, chain);
1204         }
1205
1206         error = 0;
1207 done:
1208         if (chain)
1209                 hammer2_chain_unlock(chain);
1210         if (parent)
1211                 hammer2_chain_lookup_done(parent);
1212         if (ochain)
1213                 hammer2_chain_drop(ochain);
1214
1215         return error;
1216 }
1217
1218 /*
1219  * Calculate the allocation size for the file fragment straddling EOF
1220  */
1221 int
1222 hammer2_inode_calc_alloc(hammer2_key_t filesize)
1223 {
1224         int frag = (int)filesize & HAMMER2_PBUFMASK;
1225         int radix;
1226
1227         if (frag == 0)
1228                 return(0);
1229         for (radix = HAMMER2_MINALLOCRADIX; frag > (1 << radix); ++radix)
1230                 ;
1231         return (radix);
1232 }
1233
1234 /*
1235  * Given an exclusively locked inode we consolidate its chain for hardlink
1236  * creation, adding (nlinks) to the file's link count and potentially
1237  * relocating the inode to a directory common to ip->pip and tdip.
1238  *
1239  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1240  * and returning a new locked chain.
1241  *
1242  * NOTE!  This function will also replace ip->chain.
1243  */
1244 int
1245 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1246                              hammer2_chain_t **chainp,
1247                              hammer2_inode_t *tdip, int nlinks)
1248 {
1249         hammer2_inode_data_t *ipdata;
1250         hammer2_mount_t *hmp;
1251         hammer2_inode_t *fdip;
1252         hammer2_inode_t *cdip;
1253         hammer2_chain_t *chain;
1254         hammer2_chain_t *nchain;
1255         int error;
1256
1257         hmp = tdip->hmp;
1258
1259         chain = *chainp;
1260         if (nlinks == 0 &&                      /* no hardlink needed */
1261             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1262                 return (0);
1263         }
1264         if (hammer2_hardlink_enable < 0) {      /* fake hardlinks */
1265                 return (0);
1266         }
1267
1268         if (hammer2_hardlink_enable == 0) {     /* disallow hardlinks */
1269                 hammer2_chain_unlock(chain);
1270                 *chainp = NULL;
1271                 return (ENOTSUP);
1272         }
1273
1274         /*
1275          * cdip will be returned with a ref, but not locked.
1276          */
1277         fdip = ip->pip;
1278         cdip = hammer2_inode_common_parent(fdip, tdip);
1279
1280         /*
1281          * If no change in the hardlink's target directory is required and
1282          * this is already a hardlink target, all we need to do is adjust
1283          * the link count.
1284          *
1285          * XXX The common parent is a big wiggly due to duplication from
1286          *     renames.  Compare the core (RBTREE) pointer instead of the
1287          *     ip's.
1288          */
1289         if (cdip == fdip &&
1290             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1291                 if (nlinks) {
1292                         hammer2_chain_modify(trans, &chain, 0);
1293                         chain->data->ipdata.nlinks += nlinks;
1294                 }
1295                 error = 0;
1296                 goto done;
1297         }
1298
1299         /*
1300          * We either have to move an existing hardlink target or we have
1301          * to create a fresh hardlink target.
1302          *
1303          * Hardlink targets are hidden inodes in a parent directory common
1304          * to all directory entries referencing the hardlink.
1305          */
1306         nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1307
1308         if (error == 0) {
1309                 /*
1310                  * Bump nlinks on duplicated hidden inode, repoint
1311                  * ip->chain.
1312                  */
1313                 hammer2_chain_modify(trans, &nchain, 0);
1314                 nchain->data->ipdata.nlinks += nlinks;
1315                 hammer2_inode_repoint(ip, cdip, nchain);
1316
1317                 /*
1318                  * If the old chain is not a hardlink target then replace
1319                  * it with a OBJTYPE_HARDLINK pointer.
1320                  *
1321                  * If the old chain IS a hardlink target then delete it.
1322                  */
1323                 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1324                         /*
1325                          * Replace original non-hardlink that's been dup'd
1326                          * with a special hardlink directory entry.  We must
1327                          * set the DIRECTDATA flag to prevent sub-chains
1328                          * from trying to synchronize to the inode if the
1329                          * file is extended afterwords.
1330                          */
1331                         hammer2_chain_modify(trans, &chain, 0);
1332                         ipdata = &chain->data->ipdata;
1333                         ipdata->target_type = ipdata->type;
1334                         ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1335                         ipdata->uflags = 0;
1336                         ipdata->rmajor = 0;
1337                         ipdata->rminor = 0;
1338                         ipdata->ctime = 0;
1339                         ipdata->mtime = 0;
1340                         ipdata->atime = 0;
1341                         ipdata->btime = 0;
1342                         bzero(&ipdata->uid, sizeof(ipdata->uid));
1343                         bzero(&ipdata->gid, sizeof(ipdata->gid));
1344                         ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1345                         ipdata->cap_flags = 0;
1346                         ipdata->mode = 0;
1347                         ipdata->size = 0;
1348                         ipdata->nlinks = 1;
1349                         ipdata->iparent = 0;    /* XXX */
1350                         ipdata->pfs_type = 0;
1351                         ipdata->pfs_inum = 0;
1352                         bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1353                         bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1354                         ipdata->data_quota = 0;
1355                         ipdata->data_count = 0;
1356                         ipdata->inode_quota = 0;
1357                         ipdata->inode_count = 0;
1358                         ipdata->attr_tid = 0;
1359                         ipdata->dirent_tid = 0;
1360                         bzero(&ipdata->u, sizeof(ipdata->u));
1361                         /* XXX transaction ids */
1362                 } else {
1363                         kprintf("DELETE INVISIBLE %p\n", chain);
1364                         hammer2_chain_delete(trans, chain);
1365                 }
1366
1367                 /*
1368                  * Return the new chain.
1369                  */
1370                 hammer2_chain_unlock(chain);
1371                 chain = nchain;
1372         } else {
1373                 /*
1374                  * Return an error
1375                  */
1376                 hammer2_chain_unlock(chain);
1377                 chain = NULL;
1378         }
1379
1380         /*
1381          * Cleanup, chain/nchain already dealt with.
1382          */
1383 done:
1384         *chainp = chain;
1385         hammer2_inode_drop(cdip);
1386
1387         return (error);
1388 }
1389
1390 /*
1391  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1392  * inode while (*chainp) points to the resolved (hidden hardlink
1393  * target) inode.  In this situation when nlinks is 1 we wish to
1394  * deconsolidate the hardlink, moving it back to the directory that now
1395  * represents the only remaining link.
1396  */
1397 int
1398 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1399                                hammer2_inode_t *dip,
1400                                hammer2_chain_t **chainp,
1401                                hammer2_chain_t **ochainp)
1402 {
1403         if (*ochainp == NULL)
1404                 return (0);
1405         /* XXX */
1406         return (0);
1407 }
1408
1409 /*
1410  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1411  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1412  * the *chainp and return a new locked *chainp representing the file target
1413  * (the original *chainp will be unlocked).
1414  *
1415  * When a match is found the chain representing the original HARDLINK
1416  * will be returned in *ochainp with a ref, but not locked.
1417  *
1418  * When no match is found *chainp is set to NULL and EIO is returned.
1419  * (*ochainp) will still be set to the original chain with a ref but not
1420  * locked.
1421  */
1422 int
1423 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1424                       hammer2_chain_t **ochainp)
1425 {
1426         hammer2_chain_t *chain = *chainp;
1427         hammer2_chain_t *parent;
1428         hammer2_inode_t *ip;
1429         hammer2_inode_t *pip;
1430         hammer2_key_t lhc;
1431
1432         pip = dip;
1433         hammer2_inode_ref(pip);         /* for loop */
1434         hammer2_chain_ref(chain);       /* for (*ochainp) */
1435         *ochainp = chain;
1436
1437         /*
1438          * Locate the hardlink.  pip is referenced and not locked,
1439          * ipp.
1440          *
1441          * chain is reused.
1442          */
1443         lhc = chain->data->ipdata.inum;
1444         hammer2_chain_unlock(chain);
1445         chain = NULL;
1446
1447         while ((ip = pip) != NULL) {
1448                 parent = hammer2_inode_lock_ex(ip);
1449                 hammer2_inode_drop(ip);                 /* loop */
1450                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1451                 chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1452                 hammer2_chain_lookup_done(parent);      /* discard parent */
1453                 if (chain)
1454                         break;
1455                 pip = ip->pip;          /* safe, ip held locked */
1456                 if (pip)
1457                         hammer2_inode_ref(pip);         /* loop */
1458                 hammer2_inode_unlock_ex(ip, NULL);
1459         }
1460
1461         /*
1462          * chain is locked, ip is locked.  Unlock ip, return the locked
1463          * chain.  *ipp is already set w/a ref count and not locked.
1464          *
1465          * (parent is already unlocked).
1466          */
1467         if (ip)
1468                 hammer2_inode_unlock_ex(ip, NULL);
1469         *chainp = chain;
1470         if (chain) {
1471                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1472                 /* already locked */
1473                 return (0);
1474         } else {
1475                 return (EIO);
1476         }
1477 }
1478
1479 /*
1480  * Find the directory common to both fdip and tdip, hold and return
1481  * its inode.
1482  */
1483 hammer2_inode_t *
1484 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1485 {
1486         hammer2_inode_t *scan1;
1487         hammer2_inode_t *scan2;
1488
1489         /*
1490          * We used to have a depth field but it complicated matters too
1491          * much for directory renames.  So now its ugly.  Check for
1492          * simple cases before giving up and doing it the expensive way.
1493          *
1494          * XXX need a bottom-up topology stability lock
1495          */
1496         if (fdip == tdip || fdip == tdip->pip) {
1497                 hammer2_inode_ref(fdip);
1498                 return(fdip);
1499         }
1500         if (fdip->pip == tdip) {
1501                 hammer2_inode_ref(tdip);
1502                 return(tdip);
1503         }
1504
1505         /*
1506          * XXX not MPSAFE
1507          */
1508         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1509                 scan2 = tdip;
1510                 while (scan2->pmp == tdip->pmp) {
1511                         if (scan1 == scan2) {
1512                                 hammer2_inode_ref(scan1);
1513                                 return(scan1);
1514                         }
1515                         scan2 = scan2->pip;
1516                         if (scan2 == NULL)
1517                                 break;
1518                 }
1519         }
1520         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1521               fdip, tdip);
1522         /* NOT REACHED */
1523         return(NULL);
1524 }