xref: /dflybsd-src/sys/vfs/hammer2/hammer2_inode.c (revision abd448c3b2d3508465e48d9cfdb163ef88fc242e)
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_pfsmount_t *pmp;
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 			pmp = ip->pmp;
247 			if (pmp)
248 				spin_lock(&pmp->inum_spin);
249 
250 			if (atomic_cmpset_int(&ip->refs, 1, 0)) {
251 				KKASSERT(ip->topo_cst.count == 0);
252 				if (ip->flags & HAMMER2_INODE_ONRBTREE) {
253 					atomic_clear_int(&ip->flags,
254 						     HAMMER2_INODE_ONRBTREE);
255 					RB_REMOVE(hammer2_inode_tree,
256 						  &pmp->inum_tree, ip);
257 				}
258 				if (pmp)
259 					spin_unlock(&pmp->inum_spin);
260 
261 				pip = ip->pip;
262 				ip->pip = NULL;
263 				ip->pmp = NULL;
264 
265 				/*
266 				 * Cleaning out ip->chain isn't entirely
267 				 * trivial.
268 				 */
269 				hammer2_inode_repoint(ip, NULL, NULL);
270 
271 				/*
272 				 * We have to drop pip (if non-NULL) to
273 				 * dispose of our implied reference from
274 				 * ip->pip.  We can simply loop on it.
275 				 */
276 				if (pmp) {
277 					KKASSERT((ip->flags &
278 						  HAMMER2_INODE_SROOT) == 0);
279 					kfree(ip, pmp->minode);
280 				} else {
281 					KKASSERT(ip->flags &
282 						 HAMMER2_INODE_SROOT);
283 					kfree(ip, M_HAMMER2);
284 				}
285 				ip = pip;
286 				/* continue with pip (can be NULL) */
287 			} else {
288 				if (pmp)
289 					spin_unlock(&ip->pmp->inum_spin);
290 			}
291 		} else {
292 			/*
293 			 * Non zero transition
294 			 */
295 			if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
296 				break;
297 		}
298 	}
299 }
300 
301 /*
302  * Get the vnode associated with the given inode, allocating the vnode if
303  * necessary.  The vnode will be returned exclusively locked.
304  *
305  * The caller must lock the inode (shared or exclusive).
306  *
307  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
308  * races.
309  */
310 struct vnode *
311 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
312 {
313 	hammer2_inode_data_t *ipdata;
314 	hammer2_pfsmount_t *pmp;
315 	struct vnode *vp;
316 	ccms_state_t ostate;
317 
318 	pmp = ip->pmp;
319 	KKASSERT(pmp != NULL);
320 	*errorp = 0;
321 	ipdata = &ip->chain->data->ipdata;
322 
323 	for (;;) {
324 		/*
325 		 * Attempt to reuse an existing vnode assignment.  It is
326 		 * possible to race a reclaim so the vget() may fail.  The
327 		 * inode must be unlocked during the vget() to avoid a
328 		 * deadlock against a reclaim.
329 		 */
330 		vp = ip->vp;
331 		if (vp) {
332 			/*
333 			 * Inode must be unlocked during the vget() to avoid
334 			 * possible deadlocks, but leave the ip ref intact.
335 			 *
336 			 * vnode is held to prevent destruction during the
337 			 * vget().  The vget() can still fail if we lost
338 			 * a reclaim race on the vnode.
339 			 */
340 			vhold_interlocked(vp);
341 			ostate = hammer2_inode_lock_temp_release(ip);
342 			if (vget(vp, LK_EXCLUSIVE)) {
343 				vdrop(vp);
344 				hammer2_inode_lock_temp_restore(ip, ostate);
345 				continue;
346 			}
347 			hammer2_inode_lock_temp_restore(ip, ostate);
348 			vdrop(vp);
349 			/* vp still locked and ref from vget */
350 			if (ip->vp != vp) {
351 				kprintf("hammer2: igetv race %p/%p\n",
352 					ip->vp, vp);
353 				vput(vp);
354 				continue;
355 			}
356 			*errorp = 0;
357 			break;
358 		}
359 
360 		/*
361 		 * No vnode exists, allocate a new vnode.  Beware of
362 		 * allocation races.  This function will return an
363 		 * exclusively locked and referenced vnode.
364 		 */
365 		*errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
366 		if (*errorp) {
367 			kprintf("hammer2: igetv getnewvnode failed %d\n",
368 				*errorp);
369 			vp = NULL;
370 			break;
371 		}
372 
373 		/*
374 		 * Lock the inode and check for an allocation race.
375 		 */
376 		ostate = hammer2_inode_lock_upgrade(ip);
377 		if (ip->vp != NULL) {
378 			vp->v_type = VBAD;
379 			vx_put(vp);
380 			hammer2_inode_lock_downgrade(ip, ostate);
381 			continue;
382 		}
383 
384 		switch (ipdata->type) {
385 		case HAMMER2_OBJTYPE_DIRECTORY:
386 			vp->v_type = VDIR;
387 			break;
388 		case HAMMER2_OBJTYPE_REGFILE:
389 			vp->v_type = VREG;
390 			vinitvmio(vp, ipdata->size,
391 				  HAMMER2_LBUFSIZE,
392 				  (int)ipdata->size & HAMMER2_LBUFMASK);
393 			break;
394 		case HAMMER2_OBJTYPE_SOFTLINK:
395 			/*
396 			 * XXX for now we are using the generic file_read
397 			 * and file_write code so we need a buffer cache
398 			 * association.
399 			 */
400 			vp->v_type = VLNK;
401 			vinitvmio(vp, ipdata->size,
402 				  HAMMER2_LBUFSIZE,
403 				  (int)ipdata->size & HAMMER2_LBUFMASK);
404 			break;
405 		/* XXX FIFO */
406 		default:
407 			panic("hammer2: unhandled objtype %d", ipdata->type);
408 			break;
409 		}
410 
411 		if (ip == pmp->iroot)
412 			vsetflags(vp, VROOT);
413 
414 		vp->v_data = ip;
415 		ip->vp = vp;
416 		hammer2_inode_ref(ip);		/* vp association */
417 		hammer2_inode_lock_downgrade(ip, ostate);
418 		break;
419 	}
420 
421 	/*
422 	 * Return non-NULL vp and *errorp == 0, or NULL vp and *errorp != 0.
423 	 */
424 	if (hammer2_debug & 0x0002) {
425 		kprintf("igetv vp %p refs %d aux %d\n",
426 			vp, vp->v_sysref.refcnt, vp->v_auxrefs);
427 	}
428 	return (vp);
429 }
430 
431 /*
432  * The passed-in chain must be locked and the returned inode will also be
433  * locked.  This routine typically locates or allocates the inode, assigns
434  * ip->chain (adding a ref to chain if necessary), and returns the inode.
435  *
436  * The hammer2_inode structure regulates the interface between the high level
437  * kernel VNOPS API and the filesystem backend (the chains).
438  *
439  * WARNING!  This routine sucks up the chain's lock (makes it part of the
440  *	     inode lock from the point of view of the inode lock API),
441  *	     so callers need to be careful.
442  *
443  * WARNING!  The mount code is allowed to pass dip == NULL for iroot and
444  *	     is allowed to pass pmp == NULL and dip == NULL for sroot.
445  */
446 hammer2_inode_t *
447 hammer2_inode_get(hammer2_pfsmount_t *pmp, hammer2_inode_t *dip,
448 		  hammer2_chain_t *chain)
449 {
450 	hammer2_inode_t *nip;
451 
452 	KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
453 
454 	/*
455 	 * Interlocked lookup/ref of the inode.  This code is only needed
456 	 * when looking up inodes with nlinks != 0 (TODO: optimize out
457 	 * otherwise and test for duplicates).
458 	 */
459 again:
460 	for (;;) {
461 		nip = hammer2_inode_lookup(pmp, chain->data->ipdata.inum);
462 		if (nip == NULL)
463 			break;
464 		ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
465 		if ((nip->flags & HAMMER2_INODE_ONRBTREE) == 0) { /* race */
466 			ccms_thread_unlock(&nip->topo_cst);
467 			hammer2_inode_drop(nip);
468 			continue;
469 		}
470 		if (nip->chain != chain)
471 			hammer2_inode_repoint(nip, NULL, chain);
472 
473 		/*
474 		 * Consolidated nip/nip->chain is locked (chain locked
475 		 * by caller).
476 		 */
477 		return nip;
478 	}
479 
480 	/*
481 	 * We couldn't find the inode number, create a new inode.
482 	 */
483 	if (pmp) {
484 		nip = kmalloc(sizeof(*nip), pmp->minode, M_WAITOK | M_ZERO);
485 	} else {
486 		nip = kmalloc(sizeof(*nip), M_HAMMER2, M_WAITOK | M_ZERO);
487 		nip->flags = HAMMER2_INODE_SROOT;
488 	}
489 	nip->inum = chain->data->ipdata.inum;
490 	hammer2_inode_repoint(nip, NULL, chain);
491 	nip->pip = dip;				/* can be NULL */
492 	if (dip)
493 		hammer2_inode_ref(dip);	/* ref dip for nip->pip */
494 
495 	nip->pmp = pmp;
496 
497 	/*
498 	 * ref and lock on nip gives it state compatible to after a
499 	 * hammer2_inode_lock_ex() call.
500 	 */
501 	nip->refs = 1;
502 	ccms_cst_init(&nip->topo_cst, &nip->chain);
503 	ccms_thread_lock(&nip->topo_cst, CCMS_STATE_EXCLUSIVE);
504 	/* combination of thread lock and chain lock == inode lock */
505 
506 	/*
507 	 * Attempt to add the inode.  If it fails we raced another inode
508 	 * get.  Undo all the work and try again.
509 	 */
510 	if (pmp) {
511 		spin_lock(&pmp->inum_spin);
512 		if (RB_INSERT(hammer2_inode_tree, &pmp->inum_tree, nip)) {
513 			spin_unlock(&pmp->inum_spin);
514 			ccms_thread_unlock(&nip->topo_cst);
515 			hammer2_inode_drop(nip);
516 			goto again;
517 		}
518 		atomic_set_int(&nip->flags, HAMMER2_INODE_ONRBTREE);
519 		spin_unlock(&pmp->inum_spin);
520 	}
521 
522 	return (nip);
523 }
524 
525 /*
526  * Create a new inode in the specified directory using the vattr to
527  * figure out the type of inode.
528  *
529  * If no error occurs the new inode with its chain locked is returned in
530  * *nipp, otherwise an error is returned and *nipp is set to NULL.
531  *
532  * If vap and/or cred are NULL the related fields are not set and the
533  * inode type defaults to a directory.  This is used when creating PFSs
534  * under the super-root, so the inode number is set to 1 in this case.
535  *
536  * dip is not locked on entry.
537  */
538 hammer2_inode_t *
539 hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
540 		     struct vattr *vap, struct ucred *cred,
541 		     const uint8_t *name, size_t name_len,
542 		     hammer2_chain_t **chainp, int *errorp)
543 {
544 	hammer2_inode_data_t *dipdata;
545 	hammer2_inode_data_t *nipdata;
546 	hammer2_chain_t *chain;
547 	hammer2_chain_t *parent;
548 	hammer2_inode_t *nip;
549 	hammer2_key_t lhc;
550 	int error;
551 	uid_t xuid;
552 	uuid_t dip_uid;
553 	uuid_t dip_gid;
554 	uint32_t dip_mode;
555 
556 	lhc = hammer2_dirhash(name, name_len);
557 	*errorp = 0;
558 
559 	/*
560 	 * Locate the inode or indirect block to create the new
561 	 * entry in.  At the same time check for key collisions
562 	 * and iterate until we don't get one.
563 	 *
564 	 * NOTE: hidden inodes do not have iterators.
565 	 */
566 retry:
567 	parent = hammer2_inode_lock_ex(dip);
568 	dipdata = &dip->chain->data->ipdata;
569 	dip_uid = dipdata->uid;
570 	dip_gid = dipdata->gid;
571 	dip_mode = dipdata->mode;
572 
573 	error = 0;
574 	while (error == 0) {
575 		chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
576 		if (chain == NULL)
577 			break;
578 		if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
579 			error = ENOSPC;
580 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
581 			error = ENOSPC;
582 		hammer2_chain_unlock(chain);
583 		chain = NULL;
584 		++lhc;
585 	}
586 	if (error == 0) {
587 		error = hammer2_chain_create(trans, &parent, &chain,
588 					     lhc, 0,
589 					     HAMMER2_BREF_TYPE_INODE,
590 					     HAMMER2_INODE_BYTES);
591 	}
592 
593 	/*
594 	 * Cleanup and handle retries.
595 	 */
596 	if (error == EAGAIN) {
597 		hammer2_chain_ref(parent);
598 		hammer2_inode_unlock_ex(dip, parent);
599 		hammer2_chain_wait(parent);
600 		hammer2_chain_drop(parent);
601 		goto retry;
602 	}
603 	hammer2_inode_unlock_ex(dip, parent);
604 
605 	if (error) {
606 		KKASSERT(chain == NULL);
607 		*errorp = error;
608 		return (NULL);
609 	}
610 
611 	/*
612 	 * Set up the new inode.
613 	 *
614 	 * NOTE: *_get() integrates chain's lock into the inode lock.
615 	 *
616 	 * NOTE: Only one new inode can currently be created per
617 	 *	 transaction.  If the need arises we can adjust
618 	 *	 hammer2_trans_init() to allow more.
619 	 */
620 	chain->data->ipdata.inum = trans->sync_tid;
621 	nip = hammer2_inode_get(dip->pmp, dip, chain);
622 	nipdata = &chain->data->ipdata;
623 
624 	if (vap) {
625 		KKASSERT(trans->inodes_created == 0);
626 		nipdata->type = hammer2_get_obj_type(vap->va_type);
627 		nipdata->inum = trans->sync_tid;
628 		++trans->inodes_created;
629 	} else {
630 		nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
631 		nipdata->inum = 1;
632 	}
633 	nipdata->version = HAMMER2_INODE_VERSION_ONE;
634 	hammer2_update_time(&nipdata->ctime);
635 	nipdata->mtime = nipdata->ctime;
636 	if (vap)
637 		nipdata->mode = vap->va_mode;
638 	nipdata->nlinks = 1;
639 	if (vap) {
640 		if (dip) {
641 			xuid = hammer2_to_unix_xid(&dip_uid);
642 			xuid = vop_helper_create_uid(dip->pmp->mp,
643 						     dip_mode,
644 						     xuid,
645 						     cred,
646 						     &vap->va_mode);
647 		} else {
648 			xuid = 0;
649 		}
650 		if (vap->va_vaflags & VA_UID_UUID_VALID)
651 			nipdata->uid = vap->va_uid_uuid;
652 		else if (vap->va_uid != (uid_t)VNOVAL)
653 			hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
654 		else
655 			hammer2_guid_to_uuid(&nipdata->uid, xuid);
656 
657 		if (vap->va_vaflags & VA_GID_UUID_VALID)
658 			nipdata->gid = vap->va_gid_uuid;
659 		else if (vap->va_gid != (gid_t)VNOVAL)
660 			hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
661 		else if (dip)
662 			nipdata->gid = dip_gid;
663 	}
664 
665 	/*
666 	 * Regular files and softlinks allow a small amount of data to be
667 	 * directly embedded in the inode.  This flag will be cleared if
668 	 * the size is extended past the embedded limit.
669 	 */
670 	if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
671 	    nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
672 		nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
673 	}
674 
675 	KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
676 	bcopy(name, nipdata->filename, name_len);
677 	nipdata->name_key = lhc;
678 	nipdata->name_len = name_len;
679 	*chainp = chain;
680 
681 	return (nip);
682 }
683 
684 /*
685  * chain may have been moved around by the create.
686  */
687 static
688 void
689 hammer2_chain_refactor(hammer2_chain_t **chainp)
690 {
691 	hammer2_chain_t *chain = *chainp;
692 	hammer2_chain_core_t *core;
693 
694 	core = chain->core;
695 	spin_lock(&core->cst.spin);
696 	while (hammer2_chain_refactor_test(chain, 1)) {
697 		chain = chain->next_parent;
698 		while (hammer2_chain_refactor_test(chain, 1))
699 			chain = chain->next_parent;
700 		hammer2_chain_ref(chain);
701 		spin_unlock(&core->cst.spin);
702 
703 		hammer2_chain_unlock(*chainp);
704 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
705 					  HAMMER2_RESOLVE_NOREF); /* eat ref */
706 		*chainp = chain;
707 		spin_lock(&core->cst.spin);
708 	}
709 	spin_unlock(&core->cst.spin);
710 }
711 
712 /*
713  * ochain represents the target file inode.  We need to move it to the
714  * specified common parent directory (dip) and rename it to a special
715  * invisible "0xINODENUMBER" filename.
716  *
717  * We use chain_duplicate and duplicate ochain at the new location,
718  * renaming it appropriately.  We create a temporary chain and
719  * then delete it to placemark where the duplicate will go.  Both of
720  * these use the inode number for (lhc) (the key), generating the
721  * invisible filename.
722  */
723 static
724 hammer2_chain_t *
725 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
726 			hammer2_inode_t *dip, int *errorp)
727 {
728 	hammer2_inode_data_t *nipdata;
729 	hammer2_chain_t *parent;
730 	hammer2_chain_t *ochain;
731 	hammer2_chain_t *nchain;
732 	hammer2_chain_t *tmp;
733 	hammer2_key_t lhc;
734 	hammer2_blockref_t bref;
735 
736 	ochain = *ochainp;
737 	*errorp = 0;
738 	lhc = ochain->data->ipdata.inum;
739 	KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
740 
741 	/*
742 	 * Locate the inode or indirect block to create the new
743 	 * entry in.  lhc represents the inode number so there is
744 	 * no collision iteration.
745 	 *
746 	 * There should be no key collisions with invisible inode keys.
747 	 */
748 retry:
749 	parent = hammer2_chain_lookup_init(dip->chain, 0);
750 	nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
751 	if (nchain) {
752 		kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
753 			nchain, parent, dip, dip->chain);
754 		hammer2_chain_unlock(nchain);
755 		nchain = NULL;
756 		*errorp = ENOSPC;
757 #if 1
758 		Debugger("X3");
759 #endif
760 	}
761 
762 	/*
763 	 * Create entry in common parent directory using the seek position
764 	 * calculated above.
765 	 */
766 	if (*errorp == 0) {
767 		KKASSERT(nchain == NULL);
768 		*errorp = hammer2_chain_create(trans, &parent, &nchain,
769 					       lhc, 0,
770 					       HAMMER2_BREF_TYPE_INODE,/* n/a */
771 					       HAMMER2_INODE_BYTES);   /* n/a */
772 		hammer2_chain_refactor(&ochain);
773 		*ochainp = ochain;
774 	}
775 
776 	/*
777 	 * Cleanup and handle retries.
778 	 */
779 	if (*errorp == EAGAIN) {
780 		hammer2_chain_ref(parent);
781 		hammer2_chain_lookup_done(parent);
782 		hammer2_chain_wait(parent);
783 		hammer2_chain_drop(parent);
784 		goto retry;
785 	}
786 
787 	/*
788 	 * Handle the error case
789 	 */
790 	if (*errorp) {
791 		KKASSERT(nchain == NULL);
792 		hammer2_chain_lookup_done(parent);
793 		return (NULL);
794 	}
795 
796 	/*
797 	 * Use chain as a placeholder for (lhc), delete it and replace
798 	 * it with our duplication.
799 	 *
800 	 * Gain a second lock on ochain for the duplication function to
801 	 * unlock, maintain the caller's original lock across the call.
802 	 *
803 	 * This is a bit messy.
804 	 */
805 	hammer2_chain_delete(trans, nchain);
806 	hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
807 	tmp = ochain;
808 	bref = tmp->bref;
809 	bref.key = lhc;			/* invisible dir entry key */
810 	bref.keybits = 0;
811 	hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
812 	hammer2_chain_lookup_done(parent);
813 	hammer2_chain_unlock(nchain);	/* no longer needed */
814 
815 	/*
816 	 * Now set chain to our duplicate and modify it appropriately.
817 	 *
818 	 * Directory entries are inodes but this is a hidden hardlink
819 	 * target.  The name isn't used but to ease debugging give it
820 	 * a name after its inode number.
821 	 */
822 	nchain = tmp;
823 	tmp = NULL;	/* safety */
824 
825 	hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
826 	nipdata = &nchain->data->ipdata;
827 	ksnprintf(nipdata->filename, sizeof(nipdata->filename),
828 		  "0x%016jx", (intmax_t)nipdata->inum);
829 	nipdata->name_len = strlen(nipdata->filename);
830 	nipdata->name_key = lhc;
831 
832 	return (nchain);
833 }
834 
835 /*
836  * Connect the target inode represented by (*chainp) to the media topology
837  * at (dip, name, len).
838  *
839  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
840  * entry instead of connecting (*chainp).
841  *
842  * If hlink is FALSE this function uses chain_duplicate() to make a copy
843  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
844  * by the caller in this case (e.g. rename).
845  */
846 int
847 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
848 		      hammer2_inode_t *dip, hammer2_chain_t **chainp,
849 		      const uint8_t *name, size_t name_len)
850 {
851 	hammer2_inode_data_t *ipdata;
852 	hammer2_chain_t *nchain;
853 	hammer2_chain_t *parent;
854 	hammer2_chain_t *ochain;
855 	hammer2_key_t lhc;
856 	int error;
857 
858 	ochain = *chainp;
859 
860 	/*
861 	 * Since ochain is either disconnected from the topology or represents
862 	 * a hardlink terminus which is always a parent of or equal to dip,
863 	 * we should be able to safely lock dip->chain for our setup.
864 	 */
865 	parent = hammer2_chain_lookup_init(dip->chain, 0);
866 
867 	lhc = hammer2_dirhash(name, name_len);
868 
869 	/*
870 	 * Locate the inode or indirect block to create the new
871 	 * entry in.  At the same time check for key collisions
872 	 * and iterate until we don't get one.
873 	 */
874 	error = 0;
875 	while (error == 0) {
876 		nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
877 		if (nchain == NULL)
878 			break;
879 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
880 			error = ENOSPC;
881 		hammer2_chain_unlock(nchain);
882 		nchain = NULL;
883 		++lhc;
884 	}
885 
886 	if (error == 0) {
887 		if (hlink) {
888 			/*
889 			 * Hardlink pointer needed, create totally fresh
890 			 * directory entry.
891 			 */
892 			KKASSERT(nchain == NULL);
893 			error = hammer2_chain_create(trans, &parent, &nchain,
894 						     lhc, 0,
895 						     HAMMER2_BREF_TYPE_INODE,
896 						     HAMMER2_INODE_BYTES);
897 			hammer2_chain_refactor(&ochain);
898 		} else {
899 			/*
900 			 * Reconnect the original chain and rename.  Use
901 			 * chain_duplicate().  The caller will likely delete
902 			 * or has already deleted the original chain in
903 			 * this case.
904 			 *
905 			 * NOTE: chain_duplicate() generates a new chain
906 			 *	 with CHAIN_DELETED cleared (ochain typically
907 			 *	 has it set from the file unlink).
908 			 */
909 			nchain = ochain;
910 			ochain = NULL;
911 			hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
912 			error = hammer2_chain_create(trans, &parent, &nchain,
913 						     lhc, 0,
914 						     HAMMER2_BREF_TYPE_INODE,
915 						     HAMMER2_INODE_BYTES);
916 		}
917 	}
918 
919 	/*
920 	 * Unlock stuff.
921 	 */
922 	KKASSERT(error != EAGAIN);
923 	hammer2_chain_lookup_done(parent);
924 	parent = NULL;
925 
926 	/*
927 	 * nchain should be NULL on error, leave ochain (== *chainp) alone.
928 	 */
929 	if (error) {
930 		KKASSERT(nchain == NULL);
931 		return (error);
932 	}
933 
934 	/*
935 	 * Directory entries are inodes so if the name has changed we have
936 	 * to update the inode.
937 	 *
938 	 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
939 	 * chain, the caller will access the hardlink via the actual hardlink
940 	 * target file and not the hardlink pointer entry, so we must still
941 	 * return ochain.
942 	 */
943 	if (hlink && hammer2_hardlink_enable >= 0) {
944 		/*
945 		 * Create the HARDLINK pointer.  oip represents the hardlink
946 		 * target in this situation.
947 		 *
948 		 * We will return ochain (the hardlink target).
949 		 */
950 		hammer2_chain_modify(trans, &nchain,
951 				     HAMMER2_MODIFY_ASSERTNOCOPY);
952 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
953 		ipdata = &nchain->data->ipdata;
954 		bcopy(name, ipdata->filename, name_len);
955 		ipdata->name_key = lhc;
956 		ipdata->name_len = name_len;
957 		ipdata->target_type = ochain->data->ipdata.type;
958 		ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
959 		ipdata->inum = ochain->data->ipdata.inum;
960 		ipdata->nlinks = 1;
961 		hammer2_chain_unlock(nchain);
962 		nchain = ochain;
963 		ochain = NULL;
964 	} else if (hlink && hammer2_hardlink_enable < 0) {
965 		/*
966 		 * Create a snapshot (hardlink fake mode for debugging).
967 		 * (ochain already flushed above so we can just copy the
968 		 * bref XXX).
969 		 *
970 		 * Since this is a snapshot we return nchain in the fake
971 		 * hardlink case.
972 		 */
973 		hammer2_chain_modify(trans, &nchain,
974 				     HAMMER2_MODIFY_ASSERTNOCOPY);
975 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
976 		ipdata = &nchain->data->ipdata;
977 		*ipdata = ochain->data->ipdata;
978 		bcopy(name, ipdata->filename, name_len);
979 		ipdata->name_key = lhc;
980 		ipdata->name_len = name_len;
981 		kprintf("created fake hardlink %*.*s\n",
982 			(int)name_len, (int)name_len, name);
983 	} else {
984 		/*
985 		 * nchain is a duplicate of ochain at the new location.
986 		 * We must fixup the name stored in oip.  The bref key
987 		 * has already been set up.
988 		 */
989 		hammer2_chain_modify(trans, &nchain,
990 				     HAMMER2_MODIFY_ASSERTNOCOPY);
991 		ipdata = &nchain->data->ipdata;
992 
993 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
994 		bcopy(name, ipdata->filename, name_len);
995 		ipdata->name_key = lhc;
996 		ipdata->name_len = name_len;
997 		ipdata->nlinks = 1;
998 	}
999 
1000 	/*
1001 	 * We are replacing ochain with nchain, unlock ochain.  In the
1002 	 * case where ochain is left unchanged the code above sets
1003 	 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1004 	 */
1005 	if (ochain)
1006 		hammer2_chain_unlock(ochain);
1007 	*chainp = nchain;
1008 
1009 	return (0);
1010 }
1011 
1012 /*
1013  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1014  * locked.
1015  *
1016  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1017  * and nchain is ref'd.
1018  */
1019 void
1020 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1021 		      hammer2_chain_t *nchain)
1022 {
1023 	hammer2_chain_t *ochain;
1024 	hammer2_inode_t *opip;
1025 
1026 	/*
1027 	 * Repoint ip->chain if requested.
1028 	 */
1029 	ochain = ip->chain;
1030 	ip->chain = nchain;
1031 	if (nchain)
1032 		hammer2_chain_ref(nchain);
1033 	if (ochain)
1034 		hammer2_chain_drop(ochain);
1035 
1036 	/*
1037 	 * Repoint ip->pip if requested (non-NULL pip).
1038 	 */
1039 	if (pip && ip->pip != pip) {
1040 		opip = ip->pip;
1041 		hammer2_inode_ref(pip);
1042 		ip->pip = pip;
1043 		if (opip)
1044 			hammer2_inode_drop(opip);
1045 	}
1046 }
1047 
1048 /*
1049  * Unlink the file from the specified directory inode.  The directory inode
1050  * does not need to be locked.
1051  *
1052  * isdir determines whether a directory/non-directory check should be made.
1053  * No check is made if isdir is set to -1.
1054  *
1055  * NOTE!  This function does not prevent the underlying file from still
1056  *	  being used if it has other refs (such as from an inode, or if it's
1057  *	  chain is manually held).  However, the caller is responsible for
1058  *	  fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
1059  */
1060 int
1061 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1062 		    const uint8_t *name, size_t name_len,
1063 		    int isdir, int *hlinkp)
1064 {
1065 	hammer2_inode_data_t *ipdata;
1066 	hammer2_chain_t *parent;
1067 	hammer2_chain_t *ochain;
1068 	hammer2_chain_t *chain;
1069 	hammer2_chain_t *dparent;
1070 	hammer2_chain_t *dchain;
1071 	hammer2_key_t lhc;
1072 	int error;
1073 	uint8_t type;
1074 
1075 	error = 0;
1076 	ochain = NULL;
1077 	lhc = hammer2_dirhash(name, name_len);
1078 
1079 	/*
1080 	 * Search for the filename in the directory
1081 	 */
1082 	if (hlinkp)
1083 		*hlinkp = 0;
1084 	parent = hammer2_inode_lock_ex(dip);
1085 	chain = hammer2_chain_lookup(&parent,
1086 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1087 				     0);
1088 	while (chain) {
1089 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1090 		    name_len == chain->data->ipdata.name_len &&
1091 		    bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1092 			break;
1093 		}
1094 		chain = hammer2_chain_next(&parent, chain,
1095 					   lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1096 					   0);
1097 	}
1098 	hammer2_inode_unlock_ex(dip, NULL);	/* retain parent */
1099 
1100 	/*
1101 	 * Not found or wrong type (isdir < 0 disables the type check).
1102 	 * If a hardlink pointer, type checks use the hardlink target.
1103 	 */
1104 	if (chain == NULL) {
1105 		error = ENOENT;
1106 		goto done;
1107 	}
1108 	if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1109 		if (hlinkp)
1110 			*hlinkp = 1;
1111 		type = chain->data->ipdata.target_type;
1112 	}
1113 
1114 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1115 		error = ENOTDIR;
1116 		goto done;
1117 	}
1118 	if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1119 		error = EISDIR;
1120 		goto done;
1121 	}
1122 
1123 	/*
1124 	 * Hardlink must be resolved.  We can't hold parent locked while we
1125 	 * do this or we could deadlock.
1126 	 *
1127 	 * On success chain will be adjusted to point at the hardlink target
1128 	 * and ochain will point to the hardlink pointer in the original
1129 	 * directory.  Otherwise chain remains pointing to the original.
1130 	 */
1131 	if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1132 		hammer2_chain_unlock(parent);
1133 		parent = NULL;
1134 		error = hammer2_hardlink_find(dip, &chain, &ochain);
1135 	}
1136 
1137 	/*
1138 	 * If this is a directory the directory must be empty.  However, if
1139 	 * isdir < 0 we are doing a rename and the directory does not have
1140 	 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1141 	 * and the directory does not have to be empty.
1142 	 *
1143 	 * NOTE: We check the full key range here which covers both visible
1144 	 *	 and invisible entries.  Theoretically there should be no
1145 	 *	 invisible (hardlink target) entries if there are no visible
1146 	 *	 entries.
1147 	 */
1148 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1149 		dparent = hammer2_chain_lookup_init(chain, 0);
1150 		dchain = hammer2_chain_lookup(&dparent,
1151 					      0, (hammer2_key_t)-1,
1152 					      HAMMER2_LOOKUP_NODATA);
1153 		if (dchain) {
1154 			hammer2_chain_unlock(dchain);
1155 			hammer2_chain_lookup_done(dparent);
1156 			error = ENOTEMPTY;
1157 			goto done;
1158 		}
1159 		hammer2_chain_lookup_done(dparent);
1160 		dparent = NULL;
1161 		/* dchain NULL */
1162 	}
1163 
1164 	/*
1165 	 * Ok, we can now unlink the chain.  We always decrement nlinks even
1166 	 * if the entry can be deleted in case someone has the file open and
1167 	 * does an fstat().
1168 	 *
1169 	 * The chain itself will no longer be in the on-media topology but
1170 	 * can still be flushed to the media (e.g. if an open descriptor
1171 	 * remains).  When the last vnode/ip ref goes away the chain will
1172 	 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1173 	 *
1174 	 * A non-NULL ochain indicates a hardlink.
1175 	 */
1176 	if (ochain) {
1177 		/*
1178 		 * Delete the original hardlink pointer.
1179 		 *
1180 		 * NOTE: parent from above is NULL when ochain != NULL
1181 		 *	 so we can reuse it.
1182 		 */
1183 		hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1184 		hammer2_chain_delete(trans, ochain);
1185 		hammer2_chain_unlock(ochain);
1186 
1187 		/*
1188 		 * Then decrement nlinks on hardlink target, deleting
1189 		 * the target when nlinks drops to 0.
1190 		 */
1191 		hammer2_chain_modify(trans, &chain, 0);
1192 		--chain->data->ipdata.nlinks;
1193 		if (chain->data->ipdata.nlinks == 0)
1194 			hammer2_chain_delete(trans, chain);
1195 	} else {
1196 		/*
1197 		 * Otherwise this was not a hardlink and we can just
1198 		 * remove the entry and decrement nlinks.
1199 		 *
1200 		 * NOTE: *_get() integrates chain's lock into the inode lock.
1201 		 */
1202 		hammer2_chain_modify(trans, &chain, 0);
1203 		ipdata = &chain->data->ipdata;
1204 		--ipdata->nlinks;
1205 		hammer2_chain_delete(trans, chain);
1206 	}
1207 
1208 	error = 0;
1209 done:
1210 	if (chain)
1211 		hammer2_chain_unlock(chain);
1212 	if (parent)
1213 		hammer2_chain_lookup_done(parent);
1214 	if (ochain)
1215 		hammer2_chain_drop(ochain);
1216 
1217 	return error;
1218 }
1219 
1220 /*
1221  * Given an exclusively locked inode we consolidate its chain for hardlink
1222  * creation, adding (nlinks) to the file's link count and potentially
1223  * relocating the inode to a directory common to ip->pip and tdip.
1224  *
1225  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1226  * and returning a new locked chain.
1227  *
1228  * NOTE!  This function will also replace ip->chain.
1229  */
1230 int
1231 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1232 			     hammer2_chain_t **chainp,
1233 			     hammer2_inode_t *tdip, int nlinks)
1234 {
1235 	hammer2_inode_data_t *ipdata;
1236 	hammer2_inode_t *fdip;
1237 	hammer2_inode_t *cdip;
1238 	hammer2_chain_t *chain;
1239 	hammer2_chain_t *nchain;
1240 	int error;
1241 
1242 	chain = *chainp;
1243 	if (nlinks == 0 &&			/* no hardlink needed */
1244 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1245 		return (0);
1246 	}
1247 	if (hammer2_hardlink_enable < 0) {	/* fake hardlinks */
1248 		return (0);
1249 	}
1250 
1251 	if (hammer2_hardlink_enable == 0) {	/* disallow hardlinks */
1252 		hammer2_chain_unlock(chain);
1253 		*chainp = NULL;
1254 		return (ENOTSUP);
1255 	}
1256 
1257 	/*
1258 	 * cdip will be returned with a ref, but not locked.
1259 	 */
1260 	fdip = ip->pip;
1261 	cdip = hammer2_inode_common_parent(fdip, tdip);
1262 
1263 	/*
1264 	 * If no change in the hardlink's target directory is required and
1265 	 * this is already a hardlink target, all we need to do is adjust
1266 	 * the link count.
1267 	 *
1268 	 * XXX The common parent is a big wiggly due to duplication from
1269 	 *     renames.  Compare the core (RBTREE) pointer instead of the
1270 	 *     ip's.
1271 	 */
1272 	if (cdip == fdip &&
1273 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1274 		if (nlinks) {
1275 			hammer2_chain_modify(trans, &chain, 0);
1276 			chain->data->ipdata.nlinks += nlinks;
1277 		}
1278 		error = 0;
1279 		goto done;
1280 	}
1281 
1282 	/*
1283 	 * We either have to move an existing hardlink target or we have
1284 	 * to create a fresh hardlink target.
1285 	 *
1286 	 * Hardlink targets are hidden inodes in a parent directory common
1287 	 * to all directory entries referencing the hardlink.
1288 	 */
1289 	nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1290 
1291 	if (error == 0) {
1292 		/*
1293 		 * Bump nlinks on duplicated hidden inode, repoint
1294 		 * ip->chain.
1295 		 */
1296 		hammer2_chain_modify(trans, &nchain, 0);
1297 		nchain->data->ipdata.nlinks += nlinks;
1298 		hammer2_inode_repoint(ip, cdip, nchain);
1299 
1300 		/*
1301 		 * If the old chain is not a hardlink target then replace
1302 		 * it with a OBJTYPE_HARDLINK pointer.
1303 		 *
1304 		 * If the old chain IS a hardlink target then delete it.
1305 		 */
1306 		if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1307 			/*
1308 			 * Replace original non-hardlink that's been dup'd
1309 			 * with a special hardlink directory entry.  We must
1310 			 * set the DIRECTDATA flag to prevent sub-chains
1311 			 * from trying to synchronize to the inode if the
1312 			 * file is extended afterwords.
1313 			 */
1314 			hammer2_chain_modify(trans, &chain, 0);
1315 			hammer2_chain_delete_duplicate(trans, &chain,
1316 						       HAMMER2_DELDUP_RECORE);
1317 			ipdata = &chain->data->ipdata;
1318 			ipdata->target_type = ipdata->type;
1319 			ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1320 			ipdata->uflags = 0;
1321 			ipdata->rmajor = 0;
1322 			ipdata->rminor = 0;
1323 			ipdata->ctime = 0;
1324 			ipdata->mtime = 0;
1325 			ipdata->atime = 0;
1326 			ipdata->btime = 0;
1327 			bzero(&ipdata->uid, sizeof(ipdata->uid));
1328 			bzero(&ipdata->gid, sizeof(ipdata->gid));
1329 			ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1330 			ipdata->cap_flags = 0;
1331 			ipdata->mode = 0;
1332 			ipdata->size = 0;
1333 			ipdata->nlinks = 1;
1334 			ipdata->iparent = 0;	/* XXX */
1335 			ipdata->pfs_type = 0;
1336 			ipdata->pfs_inum = 0;
1337 			bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1338 			bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1339 			ipdata->data_quota = 0;
1340 			ipdata->data_count = 0;
1341 			ipdata->inode_quota = 0;
1342 			ipdata->inode_count = 0;
1343 			ipdata->attr_tid = 0;
1344 			ipdata->dirent_tid = 0;
1345 			bzero(&ipdata->u, sizeof(ipdata->u));
1346 			/* XXX transaction ids */
1347 		} else {
1348 			hammer2_chain_delete(trans, chain);
1349 		}
1350 
1351 		/*
1352 		 * Return the new chain.
1353 		 */
1354 		hammer2_chain_unlock(chain);
1355 		chain = nchain;
1356 	} else {
1357 		/*
1358 		 * Return an error
1359 		 */
1360 		hammer2_chain_unlock(chain);
1361 		chain = NULL;
1362 	}
1363 
1364 	/*
1365 	 * Cleanup, chain/nchain already dealt with.
1366 	 */
1367 done:
1368 	*chainp = chain;
1369 	hammer2_inode_drop(cdip);
1370 
1371 	return (error);
1372 }
1373 
1374 /*
1375  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1376  * inode while (*chainp) points to the resolved (hidden hardlink
1377  * target) inode.  In this situation when nlinks is 1 we wish to
1378  * deconsolidate the hardlink, moving it back to the directory that now
1379  * represents the only remaining link.
1380  */
1381 int
1382 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1383 			       hammer2_inode_t *dip,
1384 			       hammer2_chain_t **chainp,
1385 			       hammer2_chain_t **ochainp)
1386 {
1387 	if (*ochainp == NULL)
1388 		return (0);
1389 	/* XXX */
1390 	return (0);
1391 }
1392 
1393 /*
1394  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1395  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1396  * the *chainp and return a new locked *chainp representing the file target
1397  * (the original *chainp will be unlocked).
1398  *
1399  * When a match is found the chain representing the original HARDLINK
1400  * will be returned in *ochainp with a ref, but not locked.
1401  *
1402  * When no match is found *chainp is set to NULL and EIO is returned.
1403  * (*ochainp) will still be set to the original chain with a ref but not
1404  * locked.
1405  */
1406 int
1407 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1408 		      hammer2_chain_t **ochainp)
1409 {
1410 	hammer2_chain_t *chain = *chainp;
1411 	hammer2_chain_t *parent;
1412 	hammer2_inode_t *ip;
1413 	hammer2_inode_t *pip;
1414 	hammer2_key_t lhc;
1415 
1416 	pip = dip;
1417 	hammer2_inode_ref(pip);		/* for loop */
1418 	hammer2_chain_ref(chain);	/* for (*ochainp) */
1419 	*ochainp = chain;
1420 
1421 	/*
1422 	 * Locate the hardlink.  pip is referenced and not locked,
1423 	 * ipp.
1424 	 *
1425 	 * chain is reused.
1426 	 */
1427 	lhc = chain->data->ipdata.inum;
1428 	hammer2_chain_unlock(chain);
1429 	chain = NULL;
1430 
1431 	while ((ip = pip) != NULL) {
1432 		parent = hammer2_inode_lock_ex(ip);
1433 		hammer2_inode_drop(ip);			/* loop */
1434 		KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1435 		chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1436 		hammer2_chain_lookup_done(parent);	/* discard parent */
1437 		if (chain)
1438 			break;
1439 		pip = ip->pip;		/* safe, ip held locked */
1440 		if (pip)
1441 			hammer2_inode_ref(pip);		/* loop */
1442 		hammer2_inode_unlock_ex(ip, NULL);
1443 	}
1444 
1445 	/*
1446 	 * chain is locked, ip is locked.  Unlock ip, return the locked
1447 	 * chain.  *ipp is already set w/a ref count and not locked.
1448 	 *
1449 	 * (parent is already unlocked).
1450 	 */
1451 	if (ip)
1452 		hammer2_inode_unlock_ex(ip, NULL);
1453 	*chainp = chain;
1454 	if (chain) {
1455 		KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1456 		/* already locked */
1457 		return (0);
1458 	} else {
1459 		return (EIO);
1460 	}
1461 }
1462 
1463 /*
1464  * Find the directory common to both fdip and tdip, hold and return
1465  * its inode.
1466  */
1467 hammer2_inode_t *
1468 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1469 {
1470 	hammer2_inode_t *scan1;
1471 	hammer2_inode_t *scan2;
1472 
1473 	/*
1474 	 * We used to have a depth field but it complicated matters too
1475 	 * much for directory renames.  So now its ugly.  Check for
1476 	 * simple cases before giving up and doing it the expensive way.
1477 	 *
1478 	 * XXX need a bottom-up topology stability lock
1479 	 */
1480 	if (fdip == tdip || fdip == tdip->pip) {
1481 		hammer2_inode_ref(fdip);
1482 		return(fdip);
1483 	}
1484 	if (fdip->pip == tdip) {
1485 		hammer2_inode_ref(tdip);
1486 		return(tdip);
1487 	}
1488 
1489 	/*
1490 	 * XXX not MPSAFE
1491 	 */
1492 	for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1493 		scan2 = tdip;
1494 		while (scan2->pmp == tdip->pmp) {
1495 			if (scan1 == scan2) {
1496 				hammer2_inode_ref(scan1);
1497 				return(scan1);
1498 			}
1499 			scan2 = scan2->pip;
1500 			if (scan2 == NULL)
1501 				break;
1502 		}
1503 	}
1504 	panic("hammer2_inode_common_parent: no common parent %p %p\n",
1505 	      fdip, tdip);
1506 	/* NOT REACHED */
1507 	return(NULL);
1508 }
1509