hammer2 - stabilization
[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(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                 case HAMMER2_OBJTYPE_CDEV:
397                         vp->v_type = VCHR;
398                         /* fall through */
399                 case HAMMER2_OBJTYPE_BDEV:
400                         vp->v_ops = &pmp->mp->mnt_vn_spec_ops;
401                         if (ipdata->type != HAMMER2_OBJTYPE_CDEV)
402                                 vp->v_type = VBLK;
403                         addaliasu(vp, ipdata->rmajor, ipdata->rminor);
404                         break;
405                 case HAMMER2_OBJTYPE_FIFO:
406                         vp->v_type = VFIFO;
407                         vp->v_ops = &pmp->mp->mnt_vn_fifo_ops;
408                         break;
409                 default:
410                         panic("hammer2: unhandled objtype %d", ipdata->type);
411                         break;
412                 }
413
414                 if (ip == pmp->iroot)
415                         vsetflags(vp, VROOT);
416
417                 vp->v_data = ip;
418                 ip->vp = vp;
419                 hammer2_inode_ref(ip);          /* vp association */
420                 hammer2_inode_lock_downgrade(ip, ostate);
421                 break;
422         }
423
424         /*
425          * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
426          */
427         if (hammer2_debug & 0x0002) {
428                 kprintf("igetv vp %p refs 0x%08x aux 0x%08x\n",
429                         vp, vp->v_refcnt, vp->v_auxrefs);
430         }
431         return (vp);
432 }
433
434 /*
435  * The passed-in chain must be locked and the returned inode will also be
436  * locked.  This routine typically locates or allocates the inode, assigns
437  * ip->chain (adding a ref to chain if necessary), and returns the inode.
438  *
439  * The hammer2_inode structure regulates the interface between the high level
440  * kernel VNOPS API and the filesystem backend (the chains).
441  *
442  * WARNING!  This routine sucks up the chain's lock (makes it part of the
443  *           inode lock from the point of view of the inode lock API),
444  *           so callers need to be careful.
445  *
446  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
447  *           is allowed to pass pmp == NULL and dip == NULL for sroot.
448  */
449 hammer2_inode_t *
450 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
451                   hammer2_chain_t *chain)
452 {
453         hammer2_inode_t *nip;
454
455         KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
456
457         /*
458          * Interlocked lookup/ref of the inode.  This code is only needed
459          * when looking up inodes with nlinks != 0 (TODO: optimize out
460          * otherwise and test for duplicates).
461          */
462 again:
463         for (;;) {
464                 nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
465                 if (nip == NULL)
466                         break;
467                 ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
468                 if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
469                         ccms_thread_unlock(&nip->topo_cst);
470                         hammer2_inode_drop(nip);
471                         continue;
472                 }
473                 if (nip->chain != chain)
474                         hammer2_inode_repoint(nip, NULL, chain);
475
476                 /*
477                  * Consolidated nip/nip->chain is locked (chain locked
478                  * by caller).
479                  */
480                 return nip;
481         }
482
483         /*
484          * We couldn't find the inode number, create a new inode.
485          */
486         if (pmp) {
487                 nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
488                 atomic_add_long(&pmp->inmem_inodes, 1);
489                 hammer2_chain_memory_wakeup(pmp);
490         } else {
491                 nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
492                 nip->flags = HAMMER2_INODE_SROOT;
493         }
494         nip->inum = chain->data->ipdata.inum;
495         nip->size = chain->data->ipdata.size;
496         nip->mtime = chain->data->ipdata.mtime;
497         hammer2_inode_repoint(nip, NULL, chain);
498         nip->pip = dip;                         /* can be NULL */
499         if (dip)
500                 hammer2_inode_ref(dip); /* ref dip for nip->pip */
501
502         nip->pmp = pmp;
503
504         /*
505          * ref and lock on nip gives it state compatible to after a
506          * hammer2_inode_lock_ex() call.
507          */
508         nip->refs = 1;
509         ccms_cst_init(&nip->topo_cst, &nip->chain);
510         ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
511         /* combination of thread lock and chain lock == inode lock */
512
513         /*
514          * Attempt to add the inode.  If it fails we raced another inode
515          * get.  Undo all the work and try again.
516          */
517         if (pmp) {
518                 spin_lock(&pmp->inum_spin);
519                 if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
520                         spin_unlock(&pmp->inum_spin);
521                         ccms_thread_unlock(&nip->topo_cst);
522                         hammer2_inode_drop(nip);
523                         goto again;
524                 }
525                 atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
526                 spin_unlock(&pmp->inum_spin);
527         }
528
529         return (nip);
530 }
531
532 /*
533  * Create a new inode in the specified directory using the vattr to
534  * figure out the type of inode.
535  *
536  * If no error occurs the new inode with its chain locked is returned in
537  * *nipp, otherwise an error is returned and *nipp is set to NULL.
538  *
539  * If vap and/or cred are NULL the related fields are not set and the
540  * inode type defaults to a directory.  This is used when creating PFSs
541  * under the super-root, so the inode number is set to 1 in this case.
542  *
543  * dip is not locked on entry.
544  */
545 hammer2_inode_t *
546 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
547                      struct vattr *vap, struct ucred *cred,
548                      const uint8_t *name, size_t name_len,
549                      hammer2_chain_t **chainp, int *errorp)
550 {
551         hammer2_inode_data_t *dipdata;
552         hammer2_inode_data_t *nipdata;
553         hammer2_chain_t *chain;
554         hammer2_chain_t *parent;
555         hammer2_inode_t *nip;
556         hammer2_key_t key_dummy;
557         hammer2_key_t lhc;
558         int error;
559         uid_t xuid;
560         uuid_t dip_uid;
561         uuid_t dip_gid;
562         uint32_t dip_mode;
563         uint8_t dip_algo;
564         int cache_index = -1;
565
566         lhc = hammer2_dirhash(name, name_len);
567         *errorp = 0;
568
569         /*
570          * Locate the inode or indirect block to create the new
571          * entry in.  At the same time check for key collisions
572          * and iterate until we don't get one.
573          *
574          * NOTE: hidden inodes do not have iterators.
575          */
576 retry:
577         parent = hammer2_inode_lock_ex(dip);
578         dipdata = &dip->chain->data->ipdata;
579         dip_uid = dipdata->uid;
580         dip_gid = dipdata->gid;
581         dip_mode = dipdata->mode;
582         dip_algo = dipdata->comp_algo;
583
584         error = 0;
585         while (error == 0) {
586                 chain = hammer2_chain_lookup(&parent, &key_dummy,
587                                              lhc, lhc, &cache_index, 0);
588                 if (chain == NULL)
589                         break;
590                 if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
591                         error = ENOSPC;
592                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
593                         error = ENOSPC;
594                 hammer2_chain_unlock(chain);
595                 chain = NULL;
596                 ++lhc;
597         }
598
599         if (error == 0) {
600                 error = hammer2_chain_create(trans, &parent, &chain,
601                                              lhc, 0,
602                                              HAMMER2_BREF_TYPE_INODE,
603                                              HAMMER2_INODE_BYTES);
604         }
605
606         /*
607          * Cleanup and handle retries.
608          */
609         if (error == EAGAIN) {
610                 hammer2_chain_ref(parent);
611                 hammer2_inode_unlock_ex(dip, parent);
612                 hammer2_chain_wait(parent);
613                 hammer2_chain_drop(parent);
614                 goto retry;
615         }
616         hammer2_inode_unlock_ex(dip, parent);
617
618         if (error) {
619                 KKASSERT(chain == NULL);
620                 *errorp = error;
621                 return (NULL);
622         }
623
624         /*
625          * Set up the new inode.
626          *
627          * NOTE: *_get() integrates chain's lock into the inode lock.
628          *
629          * NOTE: Only one new inode can currently be created per
630          *       transaction.  If the need arises we can adjust
631          *       hammer2_trans_init() to allow more.
632          *
633          * NOTE: nipdata will have chain's blockset data.
634          */
635         chain->data->ipdata.inum = trans->inode_tid;
636         nip = hammer2_inode_get(dip->pmp, dip, chain);
637         nipdata = &chain->data->ipdata;
638
639         if (vap) {
640                 KKASSERT(trans->inodes_created == 0);
641                 nipdata->type = hammer2_get_obj_type(vap->va_type);
642                 nipdata->inum = trans->inode_tid;
643                 ++trans->inodes_created;
644
645                 switch (nipdata->type) {
646                 case HAMMER2_OBJTYPE_CDEV:
647                 case HAMMER2_OBJTYPE_BDEV:
648                         nipdata->rmajor = vap->va_rmajor;
649                         nipdata->rminor = vap->va_rminor;
650                         break;
651                 default:
652                         break;
653                 }
654         } else {
655                 nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
656                 nipdata->inum = 1;
657         }
658         
659         /* Inherit parent's inode compression mode. */
660         nip->comp_heuristic = 0;
661         nipdata->comp_algo = dip_algo;
662         nipdata->version = HAMMER2_INODE_VERSION_ONE;
663         hammer2_update_time(&nipdata->ctime);
664         nipdata->mtime = nipdata->ctime;
665         if (vap)
666                 nipdata->mode = vap->va_mode;
667         nipdata->nlinks = 1;
668         if (vap) {
669                 if (dip && dip->pmp) {
670                         xuid = hammer2_to_unix_xid(&dip_uid);
671                         xuid = vop_helper_create_uid(dip->pmp->mp,
672                                                      dip_mode,
673                                                      xuid,
674                                                      cred,
675                                                      &vap->va_mode);
676                 } else {
677                         /* super-root has no dip and/or pmp */
678                         xuid = 0;
679                 }
680                 if (vap->va_vaflags & VA_UID_UUID_VALID)
681                         nipdata->uid = vap->va_uid_uuid;
682                 else if (vap->va_uid != (uid_t)VNOVAL)
683                         hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
684                 else
685                         hammer2_guid_to_uuid(&nipdata->uid, xuid);
686
687                 if (vap->va_vaflags & VA_GID_UUID_VALID)
688                         nipdata->gid = vap->va_gid_uuid;
689                 else if (vap->va_gid != (gid_t)VNOVAL)
690                         hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
691                 else if (dip)
692                         nipdata->gid = dip_gid;
693         }
694
695         /*
696          * Regular files and softlinks allow a small amount of data to be
697          * directly embedded in the inode.  This flag will be cleared if
698          * the size is extended past the embedded limit.
699          */
700         if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
701             nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
702                 nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
703         }
704
705         KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
706         bcopy(name, nipdata->filename, name_len);
707         nipdata->name_key = lhc;
708         nipdata->name_len = name_len;
709         *chainp = chain;
710
711         return (nip);
712 }
713
714 /*
715  * chain may have been moved around by the create.
716  */
717 void
718 hammer2_chain_refactor(hammer2_chain_t **chainp)
719 {
720         hammer2_chain_t *chain = *chainp;
721         hammer2_chain_core_t *core;
722
723         core = chain->core;
724         while (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
725                 spin_lock(&core->cst.spin);
726                 chain = TAILQ_NEXT(chain, core_entry);
727                 while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
728                         chain = TAILQ_NEXT(chain, core_entry);
729                 hammer2_chain_ref(chain);
730                 spin_unlock(&core->cst.spin);
731                 KKASSERT(chain->core == core);
732
733                 hammer2_chain_unlock(*chainp);
734                 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
735                                           HAMMER2_RESOLVE_NOREF); /* eat ref */
736                 *chainp = chain;
737         }
738 }
739
740 /*
741  * ochain represents the target file inode.  We need to move it to the
742  * specified common parent directory (dip) and rename it to a special
743  * invisible "0xINODENUMBER" filename.
744  *
745  * We use chain_duplicate and duplicate ochain at the new location,
746  * renaming it appropriately.  We create a temporary chain and
747  * then delete it to placemark where the duplicate will go.  Both of
748  * these use the inode number for (lhc) (the key), generating the
749  * invisible filename.
750  *
751  * The original ochain is deleted.
752  */
753 static
754 hammer2_chain_t *
755 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
756                         hammer2_inode_t *dip, int *errorp)
757 {
758         hammer2_inode_data_t *nipdata;
759         hammer2_chain_t *parent;
760         hammer2_chain_t *ochain;
761         hammer2_chain_t *nchain;
762         hammer2_chain_t *tmp;
763         hammer2_key_t key_dummy;
764         hammer2_key_t lhc;
765         hammer2_blockref_t bref;
766         int cache_index = -1;
767
768         ochain = *ochainp;
769         *errorp = 0;
770         lhc = ochain->data->ipdata.inum;
771         KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
772
773         /*
774          * Locate the inode or indirect block to create the new
775          * entry in.  lhc represents the inode number so there is
776          * no collision iteration.
777          *
778          * There should be no key collisions with invisible inode keys.
779          *
780          * WARNING! Must use inode_lock_ex() on dip to handle a stale
781          *          dip->chain cache.
782          */
783 retry:
784         parent = hammer2_inode_lock_ex(dip);
785         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
786         nchain = hammer2_chain_lookup(&parent, &key_dummy,
787                                       lhc, lhc, &cache_index, 0);
788         if (nchain) {
789                 kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
790                         nchain, parent, dip, dip->chain);
791                 hammer2_chain_unlock(nchain);
792                 nchain = NULL;
793                 *errorp = ENOSPC;
794 #if 0
795                 Debugger("X3");
796 #endif
797         }
798
799         /*
800          * Create entry in common parent directory using the seek position
801          * calculated above.
802          */
803         if (*errorp == 0) {
804                 KKASSERT(nchain == NULL);
805                 *errorp = hammer2_chain_create(trans, &parent, &nchain,
806                                                lhc, 0,
807                                                HAMMER2_BREF_TYPE_INODE,/* n/a */
808                                                HAMMER2_INODE_BYTES);   /* n/a */
809                 hammer2_chain_refactor(&ochain);
810                 *ochainp = ochain;
811         }
812
813         /*
814          * Cleanup and handle retries.
815          */
816         if (*errorp == EAGAIN) {
817                 hammer2_chain_ref(parent);
818                 /* hammer2_chain_lookup_done(parent); */
819                 hammer2_inode_unlock_ex(dip, parent);
820                 hammer2_chain_wait(parent);
821                 hammer2_chain_drop(parent);
822                 goto retry;
823         }
824
825         /*
826          * Handle the error case
827          */
828         if (*errorp) {
829                 KKASSERT(nchain == NULL);
830                 hammer2_inode_unlock_ex(dip, parent);
831                 /*hammer2_chain_lookup_done(parent);*/
832                 return (NULL);
833         }
834
835         /*
836          * Use nchain as a placeholder for (lhc), delete it and replace
837          * it with our duplication.
838          *
839          * Gain a second lock on ochain for the duplication function to
840          * unlock, maintain the caller's original lock across the call.
841          *
842          * This is a bit messy.
843          */
844         hammer2_chain_delete(trans, nchain, 0);
845         hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
846         tmp = ochain;
847         bref = tmp->bref;
848         bref.key = lhc;                 /* invisible dir entry key */
849         bref.keybits = 0;
850         hammer2_chain_duplicate(trans, &parent, &tmp, &bref, 0);
851         hammer2_inode_unlock_ex(dip, parent);
852         /*hammer2_chain_lookup_done(parent);*/
853         hammer2_chain_unlock(nchain);   /* no longer needed */
854
855         /*
856          * Now set nchain to our duplicate and modify it appropriately.
857          * Note that this may result in a delete-duplicate.
858          *
859          * Directory entries are inodes but this is a hidden hardlink
860          * target.  The name isn't used but to ease debugging give it
861          * a name after its inode number.
862          */
863         nchain = tmp;
864         tmp = NULL;     /* safety */
865
866         hammer2_chain_modify(trans, &nchain, 0);
867         nipdata = &nchain->data->ipdata;
868         ksnprintf(nipdata->filename, sizeof(nipdata->filename),
869                   "0x%016jx", (intmax_t)nipdata->inum);
870         nipdata->name_len = strlen(nipdata->filename);
871         nipdata->name_key = lhc;
872
873         return (nchain);
874 }
875
876 /*
877  * Connect the target inode represented by (*chainp) to the media topology
878  * at (dip, name, len).
879  *
880  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
881  * entry instead of connecting (*chainp).
882  *
883  * If hlink is FALSE this function uses chain_duplicate() to make a copy
884  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
885  * by the caller in this case (e.g. rename).
886  */
887 int
888 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
889                       hammer2_inode_t *dip, hammer2_chain_t **chainp,
890                       const uint8_t *name, size_t name_len)
891 {
892         hammer2_inode_data_t *ipdata;
893         hammer2_chain_t *nchain;
894         hammer2_chain_t *parent;
895         hammer2_chain_t *ochain;
896         hammer2_key_t key_dummy;
897         hammer2_key_t lhc;
898         int cache_index = -1;
899         int error;
900
901         ochain = *chainp;
902
903         /*
904          * Since ochain is either disconnected from the topology or represents
905          * a hardlink terminus which is always a parent of or equal to dip,
906          * we should be able to safely lock dip->chain for our setup.
907          *
908          * WARNING! Must use inode_lock_ex() on dip to handle a stale
909          *          dip->chain cache.
910          */
911         parent = hammer2_inode_lock_ex(dip);
912         /*parent = hammer2_chain_lookup_init(dip->chain, 0);*/
913
914         lhc = hammer2_dirhash(name, name_len);
915
916         /*
917          * Locate the inode or indirect block to create the new
918          * entry in.  At the same time check for key collisions
919          * and iterate until we don't get one.
920          */
921         error = 0;
922         while (error == 0) {
923                 nchain = hammer2_chain_lookup(&parent, &key_dummy,
924                                               lhc, lhc, &cache_index, 0);
925                 if (nchain == NULL)
926                         break;
927                 if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
928                         error = ENOSPC;
929                 hammer2_chain_unlock(nchain);
930                 nchain = NULL;
931                 ++lhc;
932         }
933
934         if (error == 0) {
935                 if (hlink) {
936                         /*
937                          * Hardlink pointer needed, create totally fresh
938                          * directory entry.
939                          */
940                         KKASSERT(nchain == NULL);
941                         error = hammer2_chain_create(trans, &parent, &nchain,
942                                                      lhc, 0,
943                                                      HAMMER2_BREF_TYPE_INODE,
944                                                      HAMMER2_INODE_BYTES);
945                         hammer2_chain_refactor(&ochain);
946                 } else {
947                         /*
948                          * Reconnect the original chain and rename.  Use
949                          * chain_duplicate().  The caller will likely delete
950                          * or has already deleted the original chain in
951                          * this case.
952                          *
953                          * NOTE: chain_duplicate() generates a new chain
954                          *       with CHAIN_DELETED cleared (ochain typically
955                          *       has it set from the file unlink).
956                          */
957                         nchain = ochain;
958                         ochain = NULL;
959                         hammer2_chain_duplicate(trans, NULL, &nchain, NULL, 0);
960                         error = hammer2_chain_create(trans, &parent, &nchain,
961                                                      lhc, 0,
962                                                      HAMMER2_BREF_TYPE_INODE,
963                                                      HAMMER2_INODE_BYTES);
964                 }
965         }
966
967         /*
968          * Unlock stuff.
969          */
970         KKASSERT(error != EAGAIN);
971         hammer2_inode_unlock_ex(dip, parent);
972         /*hammer2_chain_lookup_done(parent);*/
973         parent = NULL;
974
975         /*
976          * nchain should be NULL on error, leave ochain (== *chainp) alone.
977          */
978         if (error) {
979                 KKASSERT(nchain == NULL);
980                 return (error);
981         }
982
983         /*
984          * Directory entries are inodes so if the name has changed we have
985          * to update the inode.
986          *
987          * When creating an OBJTYPE_HARDLINK entry remember to unlock the
988          * chain, the caller will access the hardlink via the actual hardlink
989          * target file and not the hardlink pointer entry, so we must still
990          * return ochain.
991          */
992         if (hlink && hammer2_hardlink_enable >= 0) {
993                 /*
994                  * Create the HARDLINK pointer.  oip represents the hardlink
995                  * target in this situation.
996                  *
997                  * We will return ochain (the hardlink target).
998                  */
999                 hammer2_chain_modify(trans, &nchain,
1000                                      HAMMER2_MODIFY_ASSERTNOCOPY);
1001                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1002                 ipdata = &nchain->data->ipdata;
1003                 bcopy(name, ipdata->filename, name_len);
1004                 ipdata->name_key = lhc;
1005                 ipdata->name_len = name_len;
1006                 ipdata->target_type = ochain->data->ipdata.type;
1007                 ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1008                 ipdata->inum = ochain->data->ipdata.inum;
1009                 ipdata->nlinks = 1;
1010                 hammer2_chain_unlock(nchain);
1011                 nchain = ochain;
1012                 ochain = NULL;
1013         } else if (hlink && hammer2_hardlink_enable < 0) {
1014                 /*
1015                  * Create a snapshot (hardlink fake mode for debugging).
1016                  * (ochain already flushed above so we can just copy the
1017                  * bref XXX).
1018                  *
1019                  * Since this is a snapshot we return nchain in the fake
1020                  * hardlink case.
1021                  */
1022                 hammer2_chain_modify(trans, &nchain,
1023                                      HAMMER2_MODIFY_ASSERTNOCOPY);
1024                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1025                 ipdata = &nchain->data->ipdata;
1026                 *ipdata = ochain->data->ipdata;
1027                 bcopy(name, ipdata->filename, name_len);
1028                 ipdata->name_key = lhc;
1029                 ipdata->name_len = name_len;
1030                 atomic_clear_int(&nchain->core->flags,
1031                                  HAMMER2_CORE_COUNTEDBREFS);
1032                 kprintf("created fake hardlink %*.*s\n",
1033                         (int)name_len, (int)name_len, name);
1034         } else {
1035                 /*
1036                  * nchain is a duplicate of ochain at the new location.
1037                  * We must fixup the name stored in oip.  The bref key
1038                  * has already been set up.
1039                  */
1040                 hammer2_chain_modify(trans, &nchain,
1041                                      HAMMER2_MODIFY_ASSERTNOCOPY);
1042                 ipdata = &nchain->data->ipdata;
1043
1044                 KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1045                 bcopy(name, ipdata->filename, name_len);
1046                 ipdata->name_key = lhc;
1047                 ipdata->name_len = name_len;
1048                 ipdata->nlinks = 1;
1049         }
1050
1051         /*
1052          * We are replacing ochain with nchain, unlock ochain.  In the
1053          * case where ochain is left unchanged the code above sets
1054          * nchain to ochain and ochain to NULL, resulting in a NOP here.
1055          */
1056         if (ochain)
1057                 hammer2_chain_unlock(ochain);
1058         *chainp = nchain;
1059
1060         return (0);
1061 }
1062
1063 /*
1064  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1065  * locked.
1066  *
1067  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1068  * and nchain is ref'd.
1069  */
1070 void
1071 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1072                       hammer2_chain_t *nchain)
1073 {
1074         hammer2_chain_t *ochain;
1075         hammer2_inode_t *opip;
1076
1077         /*
1078          * Repoint ip->chain if requested.
1079          */
1080         ochain = ip->chain;
1081         ip->chain = nchain;
1082         if (nchain)
1083                 hammer2_chain_ref(nchain);
1084         if (ochain)
1085                 hammer2_chain_drop(ochain);
1086
1087         /*
1088          * Repoint ip->pip if requested (non-NULL pip).
1089          */
1090         if (pip && ip->pip != pip) {
1091                 opip = ip->pip;
1092                 hammer2_inode_ref(pip);
1093                 ip->pip = pip;
1094                 if (opip)
1095                         hammer2_inode_drop(opip);
1096         }
1097 }
1098
1099 /*
1100  * Unlink the file from the specified directory inode.  The directory inode
1101  * does not need to be locked.
1102  *
1103  * isdir determines whether a directory/non-directory check should be made.
1104  * No check is made if isdir is set to -1.
1105  *
1106  * NOTE!  The underlying file can still be active with open descriptors
1107  *        or if the chain is being manually held (e.g. for rename).
1108  *
1109  *        The caller is responsible for fixing up ip->chain if e.g. a
1110  *        rename occurs (see chain_duplicate()).
1111  */
1112 int
1113 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1114                     const uint8_t *name, size_t name_len,
1115                     int isdir, int *hlinkp)
1116 {
1117         hammer2_inode_data_t *ipdata;
1118         hammer2_chain_t *parent;
1119         hammer2_chain_t *ochain;
1120         hammer2_chain_t *chain;
1121         hammer2_chain_t *dparent;
1122         hammer2_chain_t *dchain;
1123         hammer2_key_t key_dummy;
1124         hammer2_key_t key_next;
1125         hammer2_key_t lhc;
1126         int error;
1127         int cache_index = -1;
1128         uint8_t type;
1129
1130         error = 0;
1131         ochain = NULL;
1132         lhc = hammer2_dirhash(name, name_len);
1133
1134         /*
1135          * Search for the filename in the directory
1136          */
1137         if (hlinkp)
1138                 *hlinkp = 0;
1139         parent = hammer2_inode_lock_ex(dip);
1140         chain = hammer2_chain_lookup(&parent, &key_next,
1141                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1142                                      &cache_index, 0);
1143         while (chain) {
1144                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1145                     name_len == chain->data->ipdata.name_len &&
1146                     bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1147                         break;
1148                 }
1149                 chain = hammer2_chain_next(&parent, chain, &key_next,
1150                                            key_next,
1151                                            lhc + HAMMER2_DIRHASH_LOMASK,
1152                                            &cache_index, 0);
1153         }
1154         hammer2_inode_unlock_ex(dip, NULL);     /* retain parent */
1155
1156         /*
1157          * Not found or wrong type (isdir < 0 disables the type check).
1158          * If a hardlink pointer, type checks use the hardlink target.
1159          */
1160         if (chain == NULL) {
1161                 error = ENOENT;
1162                 goto done;
1163         }
1164         if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1165                 if (hlinkp)
1166                         *hlinkp = 1;
1167                 type = chain->data->ipdata.target_type;
1168         }
1169
1170         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1171                 error = ENOTDIR;
1172                 goto done;
1173         }
1174         if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1175                 error = EISDIR;
1176                 goto done;
1177         }
1178
1179         /*
1180          * Hardlink must be resolved.  We can't hold parent locked while we
1181          * do this or we could deadlock.
1182          *
1183          * On success chain will be adjusted to point at the hardlink target
1184          * and ochain will point to the hardlink pointer in the original
1185          * directory.  Otherwise chain remains pointing to the original.
1186          */
1187         if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1188                 hammer2_chain_unlock(parent);
1189                 parent = NULL;
1190                 error = hammer2_hardlink_find(dip, &chain, &ochain);
1191         }
1192
1193         /*
1194          * If this is a directory the directory must be empty.  However, if
1195          * isdir < 0 we are doing a rename and the directory does not have
1196          * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1197          * and the directory does not have to be empty.
1198          *
1199          * NOTE: We check the full key range here which covers both visible
1200          *       and invisible entries.  Theoretically there should be no
1201          *       invisible (hardlink target) entries if there are no visible
1202          *       entries.
1203          */
1204         if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1205                 dparent = hammer2_chain_lookup_init(chain, 0);
1206                 dchain = hammer2_chain_lookup(&dparent, &key_dummy,
1207                                               0, (hammer2_key_t)-1,
1208                                               &cache_index,
1209                                               HAMMER2_LOOKUP_NODATA);
1210                 if (dchain) {
1211                         hammer2_chain_unlock(dchain);
1212                         hammer2_chain_lookup_done(dparent);
1213                         error = ENOTEMPTY;
1214                         goto done;
1215                 }
1216                 hammer2_chain_lookup_done(dparent);
1217                 dparent = NULL;
1218                 /* dchain NULL */
1219         }
1220
1221         /*
1222          * Ok, we can now unlink the chain.  We always decrement nlinks even
1223          * if the entry can be deleted in case someone has the file open and
1224          * does an fstat().
1225          *
1226          * The chain itself will no longer be in the on-media topology but
1227          * can still be flushed to the media (e.g. if an open descriptor
1228          * remains).  When the last vnode/ip ref goes away the chain will
1229          * be marked unmodified, avoiding any further (now unnecesary) I/O.
1230          *
1231          * A non-NULL ochain indicates a hardlink.
1232          */
1233         if (ochain) {
1234                 /*
1235                  * Delete the original hardlink pointer.
1236                  *
1237                  * NOTE: parent from above is NULL when ochain != NULL
1238                  *       so we can reuse it.
1239                  */
1240                 hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1241                 hammer2_chain_delete(trans, ochain, 0);
1242                 hammer2_chain_unlock(ochain);
1243
1244                 /*
1245                  * Then decrement nlinks on hardlink target, deleting
1246                  * the target when nlinks drops to 0.
1247                  */
1248                 hammer2_chain_modify(trans, &chain, 0);
1249                 --chain->data->ipdata.nlinks;
1250                 if (chain->data->ipdata.nlinks == 0)
1251                         hammer2_chain_delete(trans, chain, 0);
1252         } else {
1253                 /*
1254                  * Otherwise this was not a hardlink and we can just
1255                  * remove the entry and decrement nlinks.
1256                  *
1257                  * NOTE: *_get() integrates chain's lock into the inode lock.
1258                  */
1259                 hammer2_chain_modify(trans, &chain, 0);
1260                 ipdata = &chain->data->ipdata;
1261                 --ipdata->nlinks;
1262                 hammer2_chain_delete(trans, chain, 0);
1263         }
1264
1265         error = 0;
1266 done:
1267         if (chain)
1268                 hammer2_chain_unlock(chain);
1269         if (parent)
1270                 hammer2_chain_lookup_done(parent);
1271         if (ochain)
1272                 hammer2_chain_drop(ochain);
1273
1274         return error;
1275 }
1276
1277 /*
1278  * Given an exclusively locked inode we consolidate its chain for hardlink
1279  * creation, adding (nlinks) to the file's link count and potentially
1280  * relocating the inode to a directory common to ip->pip and tdip.
1281  *
1282  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1283  * and returning a new locked chain.
1284  *
1285  * NOTE!  This function will also replace ip->chain.
1286  */
1287 int
1288 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1289                              hammer2_chain_t **chainp,
1290                              hammer2_inode_t *tdip, int nlinks)
1291 {
1292         hammer2_inode_data_t *ipdata;
1293         hammer2_inode_t *fdip;
1294         hammer2_inode_t *cdip;
1295         hammer2_chain_t *chain;
1296         hammer2_chain_t *nchain;
1297         int error;
1298
1299         chain = *chainp;
1300         if (nlinks == 0 &&                      /* no hardlink needed */
1301             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1302                 return (0);
1303         }
1304         if (hammer2_hardlink_enable < 0) {      /* fake hardlinks */
1305                 return (0);
1306         }
1307
1308         if (hammer2_hardlink_enable == 0) {     /* disallow hardlinks */
1309                 hammer2_chain_unlock(chain);
1310                 *chainp = NULL;
1311                 return (ENOTSUP);
1312         }
1313
1314         /*
1315          * cdip will be returned with a ref, but not locked.
1316          */
1317         fdip = ip->pip;
1318         cdip = hammer2_inode_common_parent(fdip, tdip);
1319
1320         /*
1321          * If no change in the hardlink's target directory is required and
1322          * this is already a hardlink target, all we need to do is adjust
1323          * the link count.
1324          *
1325          * XXX The common parent is a big wiggly due to duplication from
1326          *     renames.  Compare the core (RBTREE) pointer instead of the
1327          *     ip's.
1328          */
1329         if (cdip == fdip &&
1330             (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1331                 if (nlinks) {
1332                         hammer2_chain_modify(trans, &chain, 0);
1333                         chain->data->ipdata.nlinks += nlinks;
1334                 }
1335                 error = 0;
1336                 goto done;
1337         }
1338
1339         /*
1340          * We either have to move an existing hardlink target or we have
1341          * to create a fresh hardlink target.
1342          *
1343          * Hardlink targets are hidden inodes in a parent directory common
1344          * to all directory entries referencing the hardlink.
1345          */
1346         nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1347
1348         if (error == 0) {
1349                 /*
1350                  * Bump nlinks on duplicated hidden inode, repoint
1351                  * ip->chain.
1352                  */
1353                 hammer2_chain_modify(trans, &nchain, 0);
1354                 nchain->data->ipdata.nlinks += nlinks;
1355                 hammer2_inode_repoint(ip, cdip, nchain);
1356
1357                 /*
1358                  * If the old chain is not a hardlink target then replace
1359                  * it with a OBJTYPE_HARDLINK pointer.
1360                  *
1361                  * If the old chain IS a hardlink target then delete it.
1362                  */
1363                 if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1364                         /*
1365                          * Replace original non-hardlink that's been dup'd
1366                          * with a special hardlink directory entry.  We must
1367                          * set the DIRECTDATA flag to prevent sub-chains
1368                          * from trying to synchronize to the inode if the
1369                          * file is extended afterwords.
1370                          *
1371                          * DELDUP_RECORE causes the new chain to NOT inherit
1372                          * the old chain's core (sub-tree).  The new chain
1373                          * will already be marked modified.
1374                          */
1375                         /*hammer2_chain_modify(trans, &chain, 0);*/
1376                         hammer2_chain_delete_duplicate(trans, &chain,
1377                                                        HAMMER2_DELDUP_RECORE);
1378                         ipdata = &chain->data->ipdata;
1379                         ipdata->target_type = ipdata->type;
1380                         ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1381                         ipdata->uflags = 0;
1382                         ipdata->rmajor = 0;
1383                         ipdata->rminor = 0;
1384                         ipdata->ctime = 0;
1385                         ipdata->mtime = 0;
1386                         ipdata->atime = 0;
1387                         ipdata->btime = 0;
1388                         bzero(&ipdata->uid, sizeof(ipdata->uid));
1389                         bzero(&ipdata->gid, sizeof(ipdata->gid));
1390                         ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1391                         ipdata->cap_flags = 0;
1392                         ipdata->mode = 0;
1393                         ipdata->size = 0;
1394                         ipdata->nlinks = 1;
1395                         ipdata->iparent = 0;    /* XXX */
1396                         ipdata->pfs_type = 0;
1397                         ipdata->pfs_inum = 0;
1398                         bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1399                         bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1400                         ipdata->data_quota = 0;
1401                         ipdata->data_count = 0;
1402                         ipdata->inode_quota = 0;
1403                         ipdata->inode_count = 0;
1404                         ipdata->attr_tid = 0;
1405                         ipdata->dirent_tid = 0;
1406                         bzero(&ipdata->u, sizeof(ipdata->u));
1407                         /* XXX transaction ids */
1408                 } else {
1409                         hammer2_chain_delete(trans, chain, 0);
1410                 }
1411
1412                 /*
1413                  * Return the new chain.
1414                  */
1415                 hammer2_chain_unlock(chain);
1416                 chain = nchain;
1417         } else {
1418                 /*
1419                  * Return an error
1420                  */
1421                 hammer2_chain_unlock(chain);
1422                 chain = NULL;
1423         }
1424
1425         /*
1426          * Cleanup, chain/nchain already dealt with.
1427          */
1428 done:
1429         *chainp = chain;
1430         hammer2_inode_drop(cdip);
1431
1432         return (error);
1433 }
1434
1435 /*
1436  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1437  * inode while (*chainp) points to the resolved (hidden hardlink
1438  * target) inode.  In this situation when nlinks is 1 we wish to
1439  * deconsolidate the hardlink, moving it back to the directory that now
1440  * represents the only remaining link.
1441  */
1442 int
1443 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1444                                hammer2_inode_t *dip,
1445                                hammer2_chain_t **chainp,
1446                                hammer2_chain_t **ochainp)
1447 {
1448         if (*ochainp == NULL)
1449                 return (0);
1450         /* XXX */
1451         return (0);
1452 }
1453
1454 /*
1455  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1456  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1457  * the *chainp and return a new locked *chainp representing the file target
1458  * (the original *chainp will be unlocked).
1459  *
1460  * When a match is found the chain representing the original HARDLINK
1461  * will be returned in *ochainp with a ref, but not locked.
1462  *
1463  * When no match is found *chainp is set to NULL and EIO is returned.
1464  * (*ochainp) will still be set to the original chain with a ref but not
1465  * locked.
1466  */
1467 int
1468 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1469                       hammer2_chain_t **ochainp)
1470 {
1471         hammer2_chain_t *chain = *chainp;
1472         hammer2_chain_t *parent;
1473         hammer2_inode_t *ip;
1474         hammer2_inode_t *pip;
1475         hammer2_key_t key_dummy;
1476         hammer2_key_t lhc;
1477         int cache_index = -1;
1478
1479         pip = dip;
1480         hammer2_inode_ref(pip);         /* for loop */
1481         hammer2_chain_ref(chain);       /* for (*ochainp) */
1482         *ochainp = chain;
1483
1484         /*
1485          * Locate the hardlink.  pip is referenced and not locked,
1486          * ipp.
1487          *
1488          * chain is reused.
1489          */
1490         lhc = chain->data->ipdata.inum;
1491         hammer2_chain_unlock(chain);
1492         chain = NULL;
1493
1494         while ((ip = pip) != NULL) {
1495                 parent = hammer2_inode_lock_ex(ip);
1496                 hammer2_inode_drop(ip);                 /* loop */
1497                 KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1498                 chain = hammer2_chain_lookup(&parent, &key_dummy,
1499                                              lhc, lhc, &cache_index, 0);
1500                 hammer2_chain_lookup_done(parent);      /* discard parent */
1501                 if (chain)
1502                         break;
1503                 pip = ip->pip;          /* safe, ip held locked */
1504                 if (pip)
1505                         hammer2_inode_ref(pip);         /* loop */
1506                 hammer2_inode_unlock_ex(ip, NULL);
1507         }
1508
1509         /*
1510          * chain is locked, ip is locked.  Unlock ip, return the locked
1511          * chain.  *ipp is already set w/a ref count and not locked.
1512          *
1513          * (parent is already unlocked).
1514          */
1515         if (ip)
1516                 hammer2_inode_unlock_ex(ip, NULL);
1517         *chainp = chain;
1518         if (chain) {
1519                 KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1520                 /* already locked */
1521                 return (0);
1522         } else {
1523                 return (EIO);
1524         }
1525 }
1526
1527 /*
1528  * Find the directory common to both fdip and tdip, hold and return
1529  * its inode.
1530  */
1531 hammer2_inode_t *
1532 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1533 {
1534         hammer2_inode_t *scan1;
1535         hammer2_inode_t *scan2;
1536
1537         /*
1538          * We used to have a depth field but it complicated matters too
1539          * much for directory renames.  So now its ugly.  Check for
1540          * simple cases before giving up and doing it the expensive way.
1541          *
1542          * XXX need a bottom-up topology stability lock
1543          */
1544         if (fdip == tdip || fdip == tdip->pip) {
1545                 hammer2_inode_ref(fdip);
1546                 return(fdip);
1547         }
1548         if (fdip->pip == tdip) {
1549                 hammer2_inode_ref(tdip);
1550                 return(tdip);
1551         }
1552
1553         /*
1554          * XXX not MPSAFE
1555          */
1556         for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1557                 scan2 = tdip;
1558                 while (scan2->pmp == tdip->pmp) {
1559                         if (scan1 == scan2) {
1560                                 hammer2_inode_ref(scan1);
1561                                 return(scan1);
1562                         }
1563                         scan2 = scan2->pip;
1564                         if (scan2 == NULL)
1565                                 break;
1566                 }
1567         }
1568         panic("hammer2_inode_common_parent: no common parent %p %p\n",
1569               fdip, tdip);
1570         /* NOT REACHED */
1571         return(NULL);
1572 }
1573
1574 /*
1575  * Synchronize the inode's frontend state with the chain state prior
1576  * to any explicit flush of the inode or any strategy write call.
1577  *
1578  * Called with a locked inode.
1579  */
1580 void
1581 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip, 
1582                     hammer2_chain_t **chainp)
1583 {
1584         hammer2_inode_data_t *ipdata;
1585         hammer2_chain_t *parent;
1586         hammer2_chain_t *chain;
1587         hammer2_key_t lbase;
1588         hammer2_key_t key_next;
1589         int cache_index;
1590
1591         ipdata = &ip->chain->data->ipdata;
1592
1593         if (ip->flags & HAMMER2_INODE_MTIME) {
1594                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1595                 atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1596                 ipdata->mtime = ip->mtime;
1597         }
1598         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1599                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1600                 ipdata->size = ip->size;
1601                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1602
1603                 /*
1604                  * We must delete any chains beyond the EOF.  The chain
1605                  * straddling the EOF will be pending in the bioq.
1606                  */
1607                 lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1608                         ~HAMMER2_PBUFMASK64;
1609                 parent = hammer2_chain_lookup_init(ip->chain, 0);
1610                 chain = hammer2_chain_lookup(&parent, &key_next,
1611                                              lbase, (hammer2_key_t)-1,
1612                                              &cache_index,
1613                                              HAMMER2_LOOKUP_NODATA);
1614                 while (chain) {
1615                         /*
1616                          * Degenerate embedded case, nothing to loop on
1617                          */
1618                         if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1619                                 hammer2_chain_unlock(chain);
1620                                 break;
1621                         }
1622                         if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1623                                 hammer2_chain_delete(trans, chain, 0);
1624                         }
1625                         chain = hammer2_chain_next(&parent, chain, &key_next,
1626                                                    key_next, (hammer2_key_t)-1,
1627                                                    &cache_index,
1628                                                    HAMMER2_LOOKUP_NODATA);
1629                 }
1630                 hammer2_chain_lookup_done(parent);
1631         } else
1632         if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1633                 ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1634                 ipdata->size = ip->size;
1635                 atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1636
1637                 /*
1638                  * When resizing larger we may not have any direct-data
1639                  * available.
1640                  */
1641                 if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1642                     ip->size > HAMMER2_EMBEDDED_BYTES) {
1643                         ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1644                         bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
1645                 }
1646         }
1647 }