xref: /dflybsd-src/sys/vfs/hammer2/hammer2_inode.c (revision 450f08dbfd98cded95c51be4079ef10f5adb3241)
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 	spin_lock(&chain->core->cst.spin);
87 	if (hammer2_chain_refactor_test(chain, 1)) {
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 	} else {
99 		spin_unlock(&chain->core->cst.spin);
100 	}
101 
102 	KKASSERT(chain != NULL);	/* for now */
103 	hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
104 
105 	/*
106 	 * Resolve duplication races
107 	 */
108 	if (hammer2_chain_refactor_test(chain, 1)) {
109 		hammer2_chain_unlock(chain);
110 		goto again;
111 	}
112 	return (chain);
113 }
114 
115 void
116 hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain)
117 {
118 	/*
119 	 * XXX this will catch parent directories too which we don't
120 	 *     really want.
121 	 */
122 	if (chain)
123 		hammer2_chain_unlock(chain);
124 	ccms_thread_unlock(&ip->topo_cst);
125 	hammer2_inode_drop(ip);
126 }
127 
128 /*
129  * NOTE: We don't combine the inode/chain lock because putting away an
130  *       inode would otherwise confuse multiple lock holders of the inode.
131  *
132  *	 Shared locks are especially sensitive to having too many shared
133  *	 lock counts (from the same thread) on certain paths which might
134  *	 need to upgrade them.  Only one count of a shared lock can be
135  *	 upgraded.
136  */
137 hammer2_chain_t *
138 hammer2_inode_lock_sh(hammer2_inode_t *ip)
139 {
140 	hammer2_chain_t *chain;
141 
142 	hammer2_inode_ref(ip);
143 again:
144 	ccms_thread_lock(&ip->topo_cst, CCMS_STATE_SHARED);
145 
146 	chain = ip->chain;
147 	KKASSERT(chain != NULL);	/* for now */
148 	hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
149 				  HAMMER2_RESOLVE_SHARED);
150 
151 	/*
152 	 * Resolve duplication races
153 	 */
154 	if (hammer2_chain_refactor_test(chain, 1)) {
155 		hammer2_chain_unlock(chain);
156 		ccms_thread_unlock(&ip->topo_cst);
157 		chain = hammer2_inode_lock_ex(ip);
158 		hammer2_inode_unlock_ex(ip, chain);
159 		goto again;
160 	}
161 	return (chain);
162 }
163 
164 void
165 hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain)
166 {
167 	if (chain)
168 		hammer2_chain_unlock(chain);
169 	ccms_thread_unlock(&ip->topo_cst);
170 	hammer2_inode_drop(ip);
171 }
172 
173 ccms_state_t
174 hammer2_inode_lock_temp_release(hammer2_inode_t *ip)
175 {
176 	return(ccms_thread_lock_temp_release(&ip->topo_cst));
177 }
178 
179 void
180 hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate)
181 {
182 	ccms_thread_lock_temp_restore(&ip->topo_cst, ostate);
183 }
184 
185 ccms_state_t
186 hammer2_inode_lock_upgrade(hammer2_inode_t *ip)
187 {
188 	return(ccms_thread_lock_upgrade(&ip->topo_cst));
189 }
190 
191 void
192 hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate)
193 {
194 	ccms_thread_lock_downgrade(&ip->topo_cst, ostate);
195 }
196 
197 /*
198  * Lookup an inode by inode number
199  */
200 hammer2_inode_t *
201 hammer2_inode_lookup(hammer2_pfsmount_t *pmp, hammer2_tid_t inum)
202 {
203 	hammer2_inode_t *ip;
204 
205 	if (pmp) {
206 		spin_lock(&pmp->inum_spin);
207 		ip = RB_LOOKUP(hammer2_inode_tree, &pmp->inum_tree, inum);
208 		if (ip)
209 			hammer2_inode_ref(ip);
210 		spin_unlock(&pmp->inum_spin);
211 	} else {
212 		ip = NULL;
213 	}
214 	return(ip);
215 }
216 
217 /*
218  * Adding a ref to an inode is only legal if the inode already has at least
219  * one ref.
220  */
221 void
222 hammer2_inode_ref(hammer2_inode_t *ip)
223 {
224 	atomic_add_int(&ip->refs, 1);
225 }
226 
227 /*
228  * Drop an inode reference, freeing the inode when the last reference goes
229  * away.
230  */
231 void
232 hammer2_inode_drop(hammer2_inode_t *ip)
233 {
234 	hammer2_pfsmount_t *pmp;
235 	hammer2_inode_t *pip;
236 	u_int refs;
237 
238 	while (ip) {
239 		refs = ip->refs;
240 		cpu_ccfence();
241 		if (refs == 1) {
242 			/*
243 			 * Transition to zero, must interlock with
244 			 * the inode inumber lookup tree (if applicable).
245 			 *
246 			 * NOTE: The super-root inode has no pmp.
247 			 */
248 			pmp = ip->pmp;
249 			if (pmp)
250 				spin_lock(&pmp->inum_spin);
251 
252 			if (atomic_cmpset_int(&ip->refs, 1, 0)) {
253 				KKASSERT(ip->topo_cst.count == 0);
254 				if (ip->flags & HAMMER2_INODE_ONRBTREE) {
255 					atomic_clear_int(&ip->flags,
256 						     HAMMER2_INODE_ONRBTREE);
257 					RB_REMOVE(hammer2_inode_tree,
258 						  &pmp->inum_tree, ip);
259 				}
260 				if (pmp)
261 					spin_unlock(&pmp->inum_spin);
262 
263 				pip = ip->pip;
264 				ip->pip = NULL;
265 				ip->pmp = NULL;
266 
267 				/*
268 				 * Cleaning out ip->chain isn't entirely
269 				 * trivial.
270 				 */
271 				hammer2_inode_repoint(ip, NULL, NULL);
272 
273 				/*
274 				 * We have to drop pip (if non-NULL) to
275 				 * dispose of our implied reference from
276 				 * ip->pip.  We can simply loop on it.
277 				 */
278 				if (pmp) {
279 					KKASSERT((ip->flags &
280 						  HAMMER2_INODE_SROOT) == 0);
281 					kfree(ip, pmp->minode);
282 					atomic_add_long(&pmp->inmem_inodes, -1);
283 				} else {
284 					KKASSERT(ip->flags &
285 						 HAMMER2_INODE_SROOT);
286 					kfree(ip, M_HAMMER2);
287 				}
288 				ip = pip;
289 				/* continue with pip (can be NULL) */
290 			} else {
291 				if (pmp)
292 					spin_unlock(&ip->pmp->inum_spin);
293 			}
294 		} else {
295 			/*
296 			 * Non zero transition
297 			 */
298 			if (atomic_cmpset_int(&ip->refs, refs, refs - 1))
299 				break;
300 		}
301 	}
302 }
303 
304 /*
305  * Get the vnode associated with the given inode, allocating the vnode if
306  * necessary.  The vnode will be returned exclusively locked.
307  *
308  * The caller must lock the inode (shared or exclusive).
309  *
310  * Great care must be taken to avoid deadlocks and vnode acquisition/reclaim
311  * races.
312  */
313 struct vnode *
314 hammer2_igetv(hammer2_inode_t *ip, int *errorp)
315 {
316 	hammer2_inode_data_t *ipdata;
317 	hammer2_pfsmount_t *pmp;
318 	struct vnode *vp;
319 	ccms_state_t ostate;
320 
321 	pmp = ip->pmp;
322 	KKASSERT(pmp != NULL);
323 	*errorp = 0;
324 	ipdata = &ip->chain->data->ipdata;
325 
326 	for (;;) {
327 		/*
328 		 * Attempt to reuse an existing vnode assignment.  It is
329 		 * possible to race a reclaim so the vget() may fail.  The
330 		 * inode must be unlocked during the vget() to avoid a
331 		 * deadlock against a reclaim.
332 		 */
333 		vp = ip->vp;
334 		if (vp) {
335 			/*
336 			 * Inode must be unlocked during the vget() to avoid
337 			 * possible deadlocks, but leave the ip ref intact.
338 			 *
339 			 * vnode is held to prevent destruction during the
340 			 * vget().  The vget() can still fail if we lost
341 			 * a reclaim race on the vnode.
342 			 */
343 			vhold_interlocked(vp);
344 			ostate = hammer2_inode_lock_temp_release(ip);
345 			if (vget(vp, LK_EXCLUSIVE)) {
346 				vdrop(vp);
347 				hammer2_inode_lock_temp_restore(ip, ostate);
348 				continue;
349 			}
350 			hammer2_inode_lock_temp_restore(ip, ostate);
351 			vdrop(vp);
352 			/* vp still locked and ref from vget */
353 			if (ip->vp != vp) {
354 				kprintf("hammer2: igetv race %p/%p\n",
355 					ip->vp, vp);
356 				vput(vp);
357 				continue;
358 			}
359 			*errorp = 0;
360 			break;
361 		}
362 
363 		/*
364 		 * No vnode exists, allocate a new vnode.  Beware of
365 		 * allocation races.  This function will return an
366 		 * exclusively locked and referenced vnode.
367 		 */
368 		*errorp = getnewvnode(VT_HAMMER2, pmp->mp, &vp, 0, 0);
369 		if (*errorp) {
370 			kprintf("hammer2: igetv getnewvnode failed %d\n",
371 				*errorp);
372 			vp = NULL;
373 			break;
374 		}
375 
376 		/*
377 		 * Lock the inode and check for an allocation race.
378 		 */
379 		ostate = hammer2_inode_lock_upgrade(ip);
380 		if (ip->vp != NULL) {
381 			vp->v_type = VBAD;
382 			vx_put(vp);
383 			hammer2_inode_lock_downgrade(ip, ostate);
384 			continue;
385 		}
386 
387 		switch (ipdata->type) {
388 		case HAMMER2_OBJTYPE_DIRECTORY:
389 			vp->v_type = VDIR;
390 			break;
391 		case HAMMER2_OBJTYPE_REGFILE:
392 			vp->v_type = VREG;
393 			vinitvmio(vp, ipdata->size,
394 				  HAMMER2_LBUFSIZE,
395 				  (int)ipdata->size & HAMMER2_LBUFMASK);
396 			break;
397 		case HAMMER2_OBJTYPE_SOFTLINK:
398 			/*
399 			 * XXX for now we are using the generic file_read
400 			 * and file_write code so we need a buffer cache
401 			 * association.
402 			 */
403 			vp->v_type = VLNK;
404 			vinitvmio(vp, ipdata->size,
405 				  HAMMER2_LBUFSIZE,
406 				  (int)ipdata->size & HAMMER2_LBUFMASK);
407 			break;
408 		/* XXX FIFO */
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 %d aux %d\n",
429 			vp, vp->v_sysref.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 lhc;
557 	int error;
558 	uid_t xuid;
559 	uuid_t dip_uid;
560 	uuid_t dip_gid;
561 	uint32_t dip_mode;
562 
563 	lhc = hammer2_dirhash(name, name_len);
564 	*errorp = 0;
565 
566 	/*
567 	 * Locate the inode or indirect block to create the new
568 	 * entry in.  At the same time check for key collisions
569 	 * and iterate until we don't get one.
570 	 *
571 	 * NOTE: hidden inodes do not have iterators.
572 	 */
573 retry:
574 	parent = hammer2_inode_lock_ex(dip);
575 	dipdata = &dip->chain->data->ipdata;
576 	dip_uid = dipdata->uid;
577 	dip_gid = dipdata->gid;
578 	dip_mode = dipdata->mode;
579 
580 	error = 0;
581 	while (error == 0) {
582 		chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
583 		if (chain == NULL)
584 			break;
585 		if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
586 			error = ENOSPC;
587 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
588 			error = ENOSPC;
589 		hammer2_chain_unlock(chain);
590 		chain = NULL;
591 		++lhc;
592 	}
593 
594 	if (error == 0) {
595 		error = hammer2_chain_create(trans, &parent, &chain,
596 					     lhc, 0,
597 					     HAMMER2_BREF_TYPE_INODE,
598 					     HAMMER2_INODE_BYTES);
599 	}
600 
601 	/*
602 	 * Cleanup and handle retries.
603 	 */
604 	if (error == EAGAIN) {
605 		hammer2_chain_ref(parent);
606 		hammer2_inode_unlock_ex(dip, parent);
607 		hammer2_chain_wait(parent);
608 		hammer2_chain_drop(parent);
609 		goto retry;
610 	}
611 	hammer2_inode_unlock_ex(dip, parent);
612 
613 	if (error) {
614 		KKASSERT(chain == NULL);
615 		*errorp = error;
616 		return (NULL);
617 	}
618 
619 	/*
620 	 * Set up the new inode.
621 	 *
622 	 * NOTE: *_get() integrates chain's lock into the inode lock.
623 	 *
624 	 * NOTE: Only one new inode can currently be created per
625 	 *	 transaction.  If the need arises we can adjust
626 	 *	 hammer2_trans_init() to allow more.
627 	 *
628 	 * NOTE: nipdata will have chain's blockset data.
629 	 */
630 	chain->data->ipdata.inum = trans->sync_tid;
631 	nip = hammer2_inode_get(dip->pmp, dip, chain);
632 	nipdata = &chain->data->ipdata;
633 
634 	if (vap) {
635 		KKASSERT(trans->inodes_created == 0);
636 		nipdata->type = hammer2_get_obj_type(vap->va_type);
637 		nipdata->inum = trans->sync_tid;
638 		++trans->inodes_created;
639 	} else {
640 		nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
641 		nipdata->inum = 1;
642 	}
643 
644 	/* Inherit parent's inode compression mode. */
645 	nip->comp_heuristic = 0;
646 	nipdata->comp_algo = dipdata->comp_algo;
647 	nipdata->version = HAMMER2_INODE_VERSION_ONE;
648 	hammer2_update_time(&nipdata->ctime);
649 	nipdata->mtime = nipdata->ctime;
650 	if (vap)
651 		nipdata->mode = vap->va_mode;
652 	nipdata->nlinks = 1;
653 	if (vap) {
654 		if (dip) {
655 			xuid = hammer2_to_unix_xid(&dip_uid);
656 			xuid = vop_helper_create_uid(dip->pmp->mp,
657 						     dip_mode,
658 						     xuid,
659 						     cred,
660 						     &vap->va_mode);
661 		} else {
662 			xuid = 0;
663 		}
664 		if (vap->va_vaflags & VA_UID_UUID_VALID)
665 			nipdata->uid = vap->va_uid_uuid;
666 		else if (vap->va_uid != (uid_t)VNOVAL)
667 			hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
668 		else
669 			hammer2_guid_to_uuid(&nipdata->uid, xuid);
670 
671 		if (vap->va_vaflags & VA_GID_UUID_VALID)
672 			nipdata->gid = vap->va_gid_uuid;
673 		else if (vap->va_gid != (gid_t)VNOVAL)
674 			hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
675 		else if (dip)
676 			nipdata->gid = dip_gid;
677 	}
678 
679 	/*
680 	 * Regular files and softlinks allow a small amount of data to be
681 	 * directly embedded in the inode.  This flag will be cleared if
682 	 * the size is extended past the embedded limit.
683 	 */
684 	if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
685 	    nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
686 		nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
687 	}
688 
689 	KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
690 	bcopy(name, nipdata->filename, name_len);
691 	nipdata->name_key = lhc;
692 	nipdata->name_len = name_len;
693 	*chainp = chain;
694 
695 	return (nip);
696 }
697 
698 /*
699  * chain may have been moved around by the create.
700  */
701 static
702 void
703 hammer2_chain_refactor(hammer2_chain_t **chainp)
704 {
705 	hammer2_chain_t *chain = *chainp;
706 	hammer2_chain_core_t *core;
707 
708 	core = chain->core;
709 	spin_lock(&core->cst.spin);
710 	while (hammer2_chain_refactor_test(chain, 1)) {
711 		chain = chain->next_parent;
712 		while (hammer2_chain_refactor_test(chain, 1))
713 			chain = chain->next_parent;
714 		hammer2_chain_ref(chain);
715 		spin_unlock(&core->cst.spin);
716 
717 		hammer2_chain_unlock(*chainp);
718 		hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS |
719 					  HAMMER2_RESOLVE_NOREF); /* eat ref */
720 		*chainp = chain;
721 		spin_lock(&core->cst.spin);
722 	}
723 	spin_unlock(&core->cst.spin);
724 }
725 
726 /*
727  * ochain represents the target file inode.  We need to move it to the
728  * specified common parent directory (dip) and rename it to a special
729  * invisible "0xINODENUMBER" filename.
730  *
731  * We use chain_duplicate and duplicate ochain at the new location,
732  * renaming it appropriately.  We create a temporary chain and
733  * then delete it to placemark where the duplicate will go.  Both of
734  * these use the inode number for (lhc) (the key), generating the
735  * invisible filename.
736  */
737 static
738 hammer2_chain_t *
739 hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
740 			hammer2_inode_t *dip, int *errorp)
741 {
742 	hammer2_inode_data_t *nipdata;
743 	hammer2_chain_t *parent;
744 	hammer2_chain_t *ochain;
745 	hammer2_chain_t *nchain;
746 	hammer2_chain_t *tmp;
747 	hammer2_key_t lhc;
748 	hammer2_blockref_t bref;
749 
750 	ochain = *ochainp;
751 	*errorp = 0;
752 	lhc = ochain->data->ipdata.inum;
753 	KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
754 
755 	/*
756 	 * Locate the inode or indirect block to create the new
757 	 * entry in.  lhc represents the inode number so there is
758 	 * no collision iteration.
759 	 *
760 	 * There should be no key collisions with invisible inode keys.
761 	 */
762 retry:
763 	parent = hammer2_chain_lookup_init(dip->chain, 0);
764 	nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
765 	if (nchain) {
766 		kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
767 			nchain, parent, dip, dip->chain);
768 		hammer2_chain_unlock(nchain);
769 		nchain = NULL;
770 		*errorp = ENOSPC;
771 #if 1
772 		Debugger("X3");
773 #endif
774 	}
775 
776 	/*
777 	 * Create entry in common parent directory using the seek position
778 	 * calculated above.
779 	 */
780 	if (*errorp == 0) {
781 		KKASSERT(nchain == NULL);
782 		*errorp = hammer2_chain_create(trans, &parent, &nchain,
783 					       lhc, 0,
784 					       HAMMER2_BREF_TYPE_INODE,/* n/a */
785 					       HAMMER2_INODE_BYTES);   /* n/a */
786 		hammer2_chain_refactor(&ochain);
787 		*ochainp = ochain;
788 	}
789 
790 	/*
791 	 * Cleanup and handle retries.
792 	 */
793 	if (*errorp == EAGAIN) {
794 		hammer2_chain_ref(parent);
795 		hammer2_chain_lookup_done(parent);
796 		hammer2_chain_wait(parent);
797 		hammer2_chain_drop(parent);
798 		goto retry;
799 	}
800 
801 	/*
802 	 * Handle the error case
803 	 */
804 	if (*errorp) {
805 		KKASSERT(nchain == NULL);
806 		hammer2_chain_lookup_done(parent);
807 		return (NULL);
808 	}
809 
810 	/*
811 	 * Use chain as a placeholder for (lhc), delete it and replace
812 	 * it with our duplication.
813 	 *
814 	 * Gain a second lock on ochain for the duplication function to
815 	 * unlock, maintain the caller's original lock across the call.
816 	 *
817 	 * This is a bit messy.
818 	 */
819 	hammer2_chain_delete(trans, nchain, HAMMER2_DELETE_WILLDUP);
820 	hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
821 	tmp = ochain;
822 	bref = tmp->bref;
823 	bref.key = lhc;			/* invisible dir entry key */
824 	bref.keybits = 0;
825 	hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
826 	hammer2_chain_lookup_done(parent);
827 	hammer2_chain_unlock(nchain);	/* no longer needed */
828 
829 	/*
830 	 * Now set chain to our duplicate and modify it appropriately.
831 	 *
832 	 * Directory entries are inodes but this is a hidden hardlink
833 	 * target.  The name isn't used but to ease debugging give it
834 	 * a name after its inode number.
835 	 */
836 	nchain = tmp;
837 	tmp = NULL;	/* safety */
838 
839 	hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
840 	nipdata = &nchain->data->ipdata;
841 	ksnprintf(nipdata->filename, sizeof(nipdata->filename),
842 		  "0x%016jx", (intmax_t)nipdata->inum);
843 	nipdata->name_len = strlen(nipdata->filename);
844 	nipdata->name_key = lhc;
845 
846 	return (nchain);
847 }
848 
849 /*
850  * Connect the target inode represented by (*chainp) to the media topology
851  * at (dip, name, len).
852  *
853  * If hlink is TRUE this function creates an OBJTYPE_HARDLINK directory
854  * entry instead of connecting (*chainp).
855  *
856  * If hlink is FALSE this function uses chain_duplicate() to make a copy
857  * if (*chainp) in the directory entry.  (*chainp) is likely to be deleted
858  * by the caller in this case (e.g. rename).
859  */
860 int
861 hammer2_inode_connect(hammer2_trans_t *trans, int hlink,
862 		      hammer2_inode_t *dip, hammer2_chain_t **chainp,
863 		      const uint8_t *name, size_t name_len)
864 {
865 	hammer2_inode_data_t *ipdata;
866 	hammer2_chain_t *nchain;
867 	hammer2_chain_t *parent;
868 	hammer2_chain_t *ochain;
869 	hammer2_key_t lhc;
870 	int error;
871 
872 	ochain = *chainp;
873 
874 	/*
875 	 * Since ochain is either disconnected from the topology or represents
876 	 * a hardlink terminus which is always a parent of or equal to dip,
877 	 * we should be able to safely lock dip->chain for our setup.
878 	 */
879 	parent = hammer2_chain_lookup_init(dip->chain, 0);
880 
881 	lhc = hammer2_dirhash(name, name_len);
882 
883 	/*
884 	 * Locate the inode or indirect block to create the new
885 	 * entry in.  At the same time check for key collisions
886 	 * and iterate until we don't get one.
887 	 */
888 	error = 0;
889 	while (error == 0) {
890 		nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
891 		if (nchain == NULL)
892 			break;
893 		if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
894 			error = ENOSPC;
895 		hammer2_chain_unlock(nchain);
896 		nchain = NULL;
897 		++lhc;
898 	}
899 
900 	if (error == 0) {
901 		if (hlink) {
902 			/*
903 			 * Hardlink pointer needed, create totally fresh
904 			 * directory entry.
905 			 */
906 			KKASSERT(nchain == NULL);
907 			error = hammer2_chain_create(trans, &parent, &nchain,
908 						     lhc, 0,
909 						     HAMMER2_BREF_TYPE_INODE,
910 						     HAMMER2_INODE_BYTES);
911 			hammer2_chain_refactor(&ochain);
912 		} else {
913 			/*
914 			 * Reconnect the original chain and rename.  Use
915 			 * chain_duplicate().  The caller will likely delete
916 			 * or has already deleted the original chain in
917 			 * this case.
918 			 *
919 			 * NOTE: chain_duplicate() generates a new chain
920 			 *	 with CHAIN_DELETED cleared (ochain typically
921 			 *	 has it set from the file unlink).
922 			 */
923 			nchain = ochain;
924 			ochain = NULL;
925 			hammer2_chain_duplicate(trans, NULL, -1, &nchain, NULL);
926 			error = hammer2_chain_create(trans, &parent, &nchain,
927 						     lhc, 0,
928 						     HAMMER2_BREF_TYPE_INODE,
929 						     HAMMER2_INODE_BYTES);
930 		}
931 	}
932 
933 	/*
934 	 * Unlock stuff.
935 	 */
936 	KKASSERT(error != EAGAIN);
937 	hammer2_chain_lookup_done(parent);
938 	parent = NULL;
939 
940 	/*
941 	 * nchain should be NULL on error, leave ochain (== *chainp) alone.
942 	 */
943 	if (error) {
944 		KKASSERT(nchain == NULL);
945 		return (error);
946 	}
947 
948 	/*
949 	 * Directory entries are inodes so if the name has changed we have
950 	 * to update the inode.
951 	 *
952 	 * When creating an OBJTYPE_HARDLINK entry remember to unlock the
953 	 * chain, the caller will access the hardlink via the actual hardlink
954 	 * target file and not the hardlink pointer entry, so we must still
955 	 * return ochain.
956 	 */
957 	if (hlink && hammer2_hardlink_enable >= 0) {
958 		/*
959 		 * Create the HARDLINK pointer.  oip represents the hardlink
960 		 * target in this situation.
961 		 *
962 		 * We will return ochain (the hardlink target).
963 		 */
964 		hammer2_chain_modify(trans, &nchain,
965 				     HAMMER2_MODIFY_ASSERTNOCOPY);
966 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
967 		ipdata = &nchain->data->ipdata;
968 		atomic_set_int(&nchain->flags, HAMMER2_CHAIN_HARDLINK);
969 		bcopy(name, ipdata->filename, name_len);
970 		ipdata->name_key = lhc;
971 		ipdata->name_len = name_len;
972 		ipdata->target_type = ochain->data->ipdata.type;
973 		ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
974 		ipdata->inum = ochain->data->ipdata.inum;
975 		ipdata->nlinks = 1;
976 		hammer2_chain_unlock(nchain);
977 		nchain = ochain;
978 		ochain = NULL;
979 	} else if (hlink && hammer2_hardlink_enable < 0) {
980 		/*
981 		 * Create a snapshot (hardlink fake mode for debugging).
982 		 * (ochain already flushed above so we can just copy the
983 		 * bref XXX).
984 		 *
985 		 * Since this is a snapshot we return nchain in the fake
986 		 * hardlink case.
987 		 */
988 		hammer2_chain_modify(trans, &nchain,
989 				     HAMMER2_MODIFY_ASSERTNOCOPY);
990 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
991 		ipdata = &nchain->data->ipdata;
992 		*ipdata = ochain->data->ipdata;
993 		bcopy(name, ipdata->filename, name_len);
994 		ipdata->name_key = lhc;
995 		ipdata->name_len = name_len;
996 		kprintf("created fake hardlink %*.*s\n",
997 			(int)name_len, (int)name_len, name);
998 	} else {
999 		/*
1000 		 * nchain is a duplicate of ochain at the new location.
1001 		 * We must fixup the name stored in oip.  The bref key
1002 		 * has already been set up.
1003 		 */
1004 		hammer2_chain_modify(trans, &nchain,
1005 				     HAMMER2_MODIFY_ASSERTNOCOPY);
1006 		ipdata = &nchain->data->ipdata;
1007 
1008 		KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
1009 		bcopy(name, ipdata->filename, name_len);
1010 		ipdata->name_key = lhc;
1011 		ipdata->name_len = name_len;
1012 		ipdata->nlinks = 1;
1013 	}
1014 
1015 	/*
1016 	 * We are replacing ochain with nchain, unlock ochain.  In the
1017 	 * case where ochain is left unchanged the code above sets
1018 	 * nchain to ochain and ochain to NULL, resulting in a NOP here.
1019 	 */
1020 	if (ochain)
1021 		hammer2_chain_unlock(ochain);
1022 	*chainp = nchain;
1023 
1024 	return (0);
1025 }
1026 
1027 /*
1028  * Repoint ip->chain to nchain.  Caller must hold the inode exclusively
1029  * locked.
1030  *
1031  * ip->chain is set to nchain.  The prior chain in ip->chain is dropped
1032  * and nchain is ref'd.
1033  */
1034 void
1035 hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
1036 		      hammer2_chain_t *nchain)
1037 {
1038 	hammer2_chain_t *ochain;
1039 	hammer2_inode_t *opip;
1040 
1041 	/*
1042 	 * Repoint ip->chain if requested.
1043 	 */
1044 	ochain = ip->chain;
1045 	ip->chain = nchain;
1046 	if (nchain)
1047 		hammer2_chain_ref(nchain);
1048 	if (ochain)
1049 		hammer2_chain_drop(ochain);
1050 
1051 	/*
1052 	 * Flag the chain for the refactor test
1053 	 */
1054 	if (nchain && nchain->data && nchain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK)
1055 		atomic_set_int(&nchain->flags, HAMMER2_CHAIN_HARDLINK);
1056 
1057 	/*
1058 	 * Repoint ip->pip if requested (non-NULL pip).
1059 	 */
1060 	if (pip && ip->pip != pip) {
1061 		opip = ip->pip;
1062 		hammer2_inode_ref(pip);
1063 		ip->pip = pip;
1064 		if (opip)
1065 			hammer2_inode_drop(opip);
1066 	}
1067 }
1068 
1069 /*
1070  * Unlink the file from the specified directory inode.  The directory inode
1071  * does not need to be locked.
1072  *
1073  * isdir determines whether a directory/non-directory check should be made.
1074  * No check is made if isdir is set to -1.
1075  *
1076  * NOTE!  The underlying file can still be active with open descriptors
1077  *	  or if the chain is being manually held (e.g. for rename).
1078  *
1079  *	  The caller is responsible for fixing up ip->chain if e.g. a
1080  *	  rename occurs (see chain_duplicate()).
1081  */
1082 int
1083 hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
1084 		    const uint8_t *name, size_t name_len,
1085 		    int isdir, int *hlinkp)
1086 {
1087 	hammer2_inode_data_t *ipdata;
1088 	hammer2_chain_t *parent;
1089 	hammer2_chain_t *ochain;
1090 	hammer2_chain_t *chain;
1091 	hammer2_chain_t *dparent;
1092 	hammer2_chain_t *dchain;
1093 	hammer2_key_t lhc;
1094 	int error;
1095 	uint8_t type;
1096 
1097 	error = 0;
1098 	ochain = NULL;
1099 	lhc = hammer2_dirhash(name, name_len);
1100 
1101 	/*
1102 	 * Search for the filename in the directory
1103 	 */
1104 	if (hlinkp)
1105 		*hlinkp = 0;
1106 	parent = hammer2_inode_lock_ex(dip);
1107 	chain = hammer2_chain_lookup(&parent,
1108 				     lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1109 				     0);
1110 	while (chain) {
1111 		if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1112 		    name_len == chain->data->ipdata.name_len &&
1113 		    bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
1114 			break;
1115 		}
1116 		chain = hammer2_chain_next(&parent, chain,
1117 					   lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1118 					   0);
1119 	}
1120 	hammer2_inode_unlock_ex(dip, NULL);	/* retain parent */
1121 
1122 	/*
1123 	 * Not found or wrong type (isdir < 0 disables the type check).
1124 	 * If a hardlink pointer, type checks use the hardlink target.
1125 	 */
1126 	if (chain == NULL) {
1127 		error = ENOENT;
1128 		goto done;
1129 	}
1130 	if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
1131 		if (hlinkp)
1132 			*hlinkp = 1;
1133 		type = chain->data->ipdata.target_type;
1134 	}
1135 
1136 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
1137 		error = ENOTDIR;
1138 		goto done;
1139 	}
1140 	if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
1141 		error = EISDIR;
1142 		goto done;
1143 	}
1144 
1145 	/*
1146 	 * Hardlink must be resolved.  We can't hold parent locked while we
1147 	 * do this or we could deadlock.
1148 	 *
1149 	 * On success chain will be adjusted to point at the hardlink target
1150 	 * and ochain will point to the hardlink pointer in the original
1151 	 * directory.  Otherwise chain remains pointing to the original.
1152 	 */
1153 	if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
1154 		hammer2_chain_unlock(parent);
1155 		parent = NULL;
1156 		error = hammer2_hardlink_find(dip, &chain, &ochain);
1157 	}
1158 
1159 	/*
1160 	 * If this is a directory the directory must be empty.  However, if
1161 	 * isdir < 0 we are doing a rename and the directory does not have
1162 	 * to be empty, and if isdir > 1 we are deleting a PFS/snapshot
1163 	 * and the directory does not have to be empty.
1164 	 *
1165 	 * NOTE: We check the full key range here which covers both visible
1166 	 *	 and invisible entries.  Theoretically there should be no
1167 	 *	 invisible (hardlink target) entries if there are no visible
1168 	 *	 entries.
1169 	 */
1170 	if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
1171 		dparent = hammer2_chain_lookup_init(chain, 0);
1172 		dchain = hammer2_chain_lookup(&dparent,
1173 					      0, (hammer2_key_t)-1,
1174 					      HAMMER2_LOOKUP_NODATA);
1175 		if (dchain) {
1176 			hammer2_chain_unlock(dchain);
1177 			hammer2_chain_lookup_done(dparent);
1178 			error = ENOTEMPTY;
1179 			goto done;
1180 		}
1181 		hammer2_chain_lookup_done(dparent);
1182 		dparent = NULL;
1183 		/* dchain NULL */
1184 	}
1185 
1186 	/*
1187 	 * Ok, we can now unlink the chain.  We always decrement nlinks even
1188 	 * if the entry can be deleted in case someone has the file open and
1189 	 * does an fstat().
1190 	 *
1191 	 * The chain itself will no longer be in the on-media topology but
1192 	 * can still be flushed to the media (e.g. if an open descriptor
1193 	 * remains).  When the last vnode/ip ref goes away the chain will
1194 	 * be marked unmodified, avoiding any further (now unnecesary) I/O.
1195 	 *
1196 	 * A non-NULL ochain indicates a hardlink.
1197 	 */
1198 	if (ochain) {
1199 		/*
1200 		 * Delete the original hardlink pointer.
1201 		 *
1202 		 * NOTE: parent from above is NULL when ochain != NULL
1203 		 *	 so we can reuse it.
1204 		 */
1205 		hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
1206 		hammer2_chain_delete(trans, ochain, 0);
1207 		hammer2_chain_unlock(ochain);
1208 
1209 		/*
1210 		 * Then decrement nlinks on hardlink target, deleting
1211 		 * the target when nlinks drops to 0.
1212 		 */
1213 		hammer2_chain_modify(trans, &chain, 0);
1214 		--chain->data->ipdata.nlinks;
1215 		if (chain->data->ipdata.nlinks == 0)
1216 			hammer2_chain_delete(trans, chain, 0);
1217 	} else {
1218 		/*
1219 		 * Otherwise this was not a hardlink and we can just
1220 		 * remove the entry and decrement nlinks.
1221 		 *
1222 		 * NOTE: *_get() integrates chain's lock into the inode lock.
1223 		 */
1224 		hammer2_chain_modify(trans, &chain, 0);
1225 		ipdata = &chain->data->ipdata;
1226 		--ipdata->nlinks;
1227 		hammer2_chain_delete(trans, chain, 0);
1228 	}
1229 
1230 	error = 0;
1231 done:
1232 	if (chain)
1233 		hammer2_chain_unlock(chain);
1234 	if (parent)
1235 		hammer2_chain_lookup_done(parent);
1236 	if (ochain)
1237 		hammer2_chain_drop(ochain);
1238 
1239 	return error;
1240 }
1241 
1242 /*
1243  * Given an exclusively locked inode we consolidate its chain for hardlink
1244  * creation, adding (nlinks) to the file's link count and potentially
1245  * relocating the inode to a directory common to ip->pip and tdip.
1246  *
1247  * Replaces (*chainp) if consolidation occurred, unlocking the old chain
1248  * and returning a new locked chain.
1249  *
1250  * NOTE!  This function will also replace ip->chain.
1251  */
1252 int
1253 hammer2_hardlink_consolidate(hammer2_trans_t *trans, hammer2_inode_t *ip,
1254 			     hammer2_chain_t **chainp,
1255 			     hammer2_inode_t *tdip, int nlinks)
1256 {
1257 	hammer2_inode_data_t *ipdata;
1258 	hammer2_inode_t *fdip;
1259 	hammer2_inode_t *cdip;
1260 	hammer2_chain_t *chain;
1261 	hammer2_chain_t *nchain;
1262 	int error;
1263 
1264 	chain = *chainp;
1265 	if (nlinks == 0 &&			/* no hardlink needed */
1266 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE)) {
1267 		return (0);
1268 	}
1269 	if (hammer2_hardlink_enable < 0) {	/* fake hardlinks */
1270 		return (0);
1271 	}
1272 
1273 	if (hammer2_hardlink_enable == 0) {	/* disallow hardlinks */
1274 		hammer2_chain_unlock(chain);
1275 		*chainp = NULL;
1276 		return (ENOTSUP);
1277 	}
1278 
1279 	/*
1280 	 * cdip will be returned with a ref, but not locked.
1281 	 */
1282 	fdip = ip->pip;
1283 	cdip = hammer2_inode_common_parent(fdip, tdip);
1284 
1285 	/*
1286 	 * If no change in the hardlink's target directory is required and
1287 	 * this is already a hardlink target, all we need to do is adjust
1288 	 * the link count.
1289 	 *
1290 	 * XXX The common parent is a big wiggly due to duplication from
1291 	 *     renames.  Compare the core (RBTREE) pointer instead of the
1292 	 *     ip's.
1293 	 */
1294 	if (cdip == fdip &&
1295 	    (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) == 0) {
1296 		if (nlinks) {
1297 			hammer2_chain_modify(trans, &chain, 0);
1298 			chain->data->ipdata.nlinks += nlinks;
1299 		}
1300 		error = 0;
1301 		goto done;
1302 	}
1303 
1304 	/*
1305 	 * We either have to move an existing hardlink target or we have
1306 	 * to create a fresh hardlink target.
1307 	 *
1308 	 * Hardlink targets are hidden inodes in a parent directory common
1309 	 * to all directory entries referencing the hardlink.
1310 	 */
1311 	nchain = hammer2_hardlink_shiftup(trans, &chain, cdip, &error);
1312 
1313 	if (error == 0) {
1314 		/*
1315 		 * Bump nlinks on duplicated hidden inode, repoint
1316 		 * ip->chain.
1317 		 */
1318 		hammer2_chain_modify(trans, &nchain, 0);
1319 		nchain->data->ipdata.nlinks += nlinks;
1320 		hammer2_inode_repoint(ip, cdip, nchain);
1321 
1322 		/*
1323 		 * If the old chain is not a hardlink target then replace
1324 		 * it with a OBJTYPE_HARDLINK pointer.
1325 		 *
1326 		 * If the old chain IS a hardlink target then delete it.
1327 		 */
1328 		if (chain->data->ipdata.name_key & HAMMER2_DIRHASH_VISIBLE) {
1329 			/*
1330 			 * Replace original non-hardlink that's been dup'd
1331 			 * with a special hardlink directory entry.  We must
1332 			 * set the DIRECTDATA flag to prevent sub-chains
1333 			 * from trying to synchronize to the inode if the
1334 			 * file is extended afterwords.
1335 			 */
1336 			hammer2_chain_modify(trans, &chain, 0);
1337 			hammer2_chain_delete_duplicate(trans, &chain,
1338 						       HAMMER2_DELDUP_RECORE);
1339 			atomic_set_int(&chain->flags, HAMMER2_CHAIN_HARDLINK);
1340 			ipdata = &chain->data->ipdata;
1341 			ipdata->target_type = ipdata->type;
1342 			ipdata->type = HAMMER2_OBJTYPE_HARDLINK;
1343 			ipdata->uflags = 0;
1344 			ipdata->rmajor = 0;
1345 			ipdata->rminor = 0;
1346 			ipdata->ctime = 0;
1347 			ipdata->mtime = 0;
1348 			ipdata->atime = 0;
1349 			ipdata->btime = 0;
1350 			bzero(&ipdata->uid, sizeof(ipdata->uid));
1351 			bzero(&ipdata->gid, sizeof(ipdata->gid));
1352 			ipdata->op_flags = HAMMER2_OPFLAG_DIRECTDATA;
1353 			ipdata->cap_flags = 0;
1354 			ipdata->mode = 0;
1355 			ipdata->size = 0;
1356 			ipdata->nlinks = 1;
1357 			ipdata->iparent = 0;	/* XXX */
1358 			ipdata->pfs_type = 0;
1359 			ipdata->pfs_inum = 0;
1360 			bzero(&ipdata->pfs_clid, sizeof(ipdata->pfs_clid));
1361 			bzero(&ipdata->pfs_fsid, sizeof(ipdata->pfs_fsid));
1362 			ipdata->data_quota = 0;
1363 			ipdata->data_count = 0;
1364 			ipdata->inode_quota = 0;
1365 			ipdata->inode_count = 0;
1366 			ipdata->attr_tid = 0;
1367 			ipdata->dirent_tid = 0;
1368 			bzero(&ipdata->u, sizeof(ipdata->u));
1369 			/* XXX transaction ids */
1370 		} else {
1371 			hammer2_chain_delete(trans, chain, 0);
1372 		}
1373 
1374 		/*
1375 		 * Return the new chain.
1376 		 */
1377 		hammer2_chain_unlock(chain);
1378 		chain = nchain;
1379 	} else {
1380 		/*
1381 		 * Return an error
1382 		 */
1383 		hammer2_chain_unlock(chain);
1384 		chain = NULL;
1385 	}
1386 
1387 	/*
1388 	 * Cleanup, chain/nchain already dealt with.
1389 	 */
1390 done:
1391 	*chainp = chain;
1392 	hammer2_inode_drop(cdip);
1393 
1394 	return (error);
1395 }
1396 
1397 /*
1398  * If (*ochainp) is non-NULL it points to the forward OBJTYPE_HARDLINK
1399  * inode while (*chainp) points to the resolved (hidden hardlink
1400  * target) inode.  In this situation when nlinks is 1 we wish to
1401  * deconsolidate the hardlink, moving it back to the directory that now
1402  * represents the only remaining link.
1403  */
1404 int
1405 hammer2_hardlink_deconsolidate(hammer2_trans_t *trans,
1406 			       hammer2_inode_t *dip,
1407 			       hammer2_chain_t **chainp,
1408 			       hammer2_chain_t **ochainp)
1409 {
1410 	if (*ochainp == NULL)
1411 		return (0);
1412 	/* XXX */
1413 	return (0);
1414 }
1415 
1416 /*
1417  * The caller presents a locked *chainp pointing to a HAMMER2_BREF_TYPE_INODE
1418  * with an obj_type of HAMMER2_OBJTYPE_HARDLINK.  This routine will gobble
1419  * the *chainp and return a new locked *chainp representing the file target
1420  * (the original *chainp will be unlocked).
1421  *
1422  * When a match is found the chain representing the original HARDLINK
1423  * will be returned in *ochainp with a ref, but not locked.
1424  *
1425  * When no match is found *chainp is set to NULL and EIO is returned.
1426  * (*ochainp) will still be set to the original chain with a ref but not
1427  * locked.
1428  */
1429 int
1430 hammer2_hardlink_find(hammer2_inode_t *dip, hammer2_chain_t **chainp,
1431 		      hammer2_chain_t **ochainp)
1432 {
1433 	hammer2_chain_t *chain = *chainp;
1434 	hammer2_chain_t *parent;
1435 	hammer2_inode_t *ip;
1436 	hammer2_inode_t *pip;
1437 	hammer2_key_t lhc;
1438 
1439 	pip = dip;
1440 	hammer2_inode_ref(pip);		/* for loop */
1441 	hammer2_chain_ref(chain);	/* for (*ochainp) */
1442 	*ochainp = chain;
1443 
1444 	/*
1445 	 * Locate the hardlink.  pip is referenced and not locked,
1446 	 * ipp.
1447 	 *
1448 	 * chain is reused.
1449 	 */
1450 	lhc = chain->data->ipdata.inum;
1451 	hammer2_chain_unlock(chain);
1452 	chain = NULL;
1453 
1454 	while ((ip = pip) != NULL) {
1455 		parent = hammer2_inode_lock_ex(ip);
1456 		hammer2_inode_drop(ip);			/* loop */
1457 		KKASSERT(parent->bref.type == HAMMER2_BREF_TYPE_INODE);
1458 		chain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
1459 		hammer2_chain_lookup_done(parent);	/* discard parent */
1460 		if (chain)
1461 			break;
1462 		pip = ip->pip;		/* safe, ip held locked */
1463 		if (pip)
1464 			hammer2_inode_ref(pip);		/* loop */
1465 		hammer2_inode_unlock_ex(ip, NULL);
1466 	}
1467 
1468 	/*
1469 	 * chain is locked, ip is locked.  Unlock ip, return the locked
1470 	 * chain.  *ipp is already set w/a ref count and not locked.
1471 	 *
1472 	 * (parent is already unlocked).
1473 	 */
1474 	if (ip)
1475 		hammer2_inode_unlock_ex(ip, NULL);
1476 	*chainp = chain;
1477 	if (chain) {
1478 		KKASSERT(chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1479 		/* already locked */
1480 		return (0);
1481 	} else {
1482 		return (EIO);
1483 	}
1484 }
1485 
1486 /*
1487  * Find the directory common to both fdip and tdip, hold and return
1488  * its inode.
1489  */
1490 hammer2_inode_t *
1491 hammer2_inode_common_parent(hammer2_inode_t *fdip, hammer2_inode_t *tdip)
1492 {
1493 	hammer2_inode_t *scan1;
1494 	hammer2_inode_t *scan2;
1495 
1496 	/*
1497 	 * We used to have a depth field but it complicated matters too
1498 	 * much for directory renames.  So now its ugly.  Check for
1499 	 * simple cases before giving up and doing it the expensive way.
1500 	 *
1501 	 * XXX need a bottom-up topology stability lock
1502 	 */
1503 	if (fdip == tdip || fdip == tdip->pip) {
1504 		hammer2_inode_ref(fdip);
1505 		return(fdip);
1506 	}
1507 	if (fdip->pip == tdip) {
1508 		hammer2_inode_ref(tdip);
1509 		return(tdip);
1510 	}
1511 
1512 	/*
1513 	 * XXX not MPSAFE
1514 	 */
1515 	for (scan1 = fdip; scan1->pmp == fdip->pmp; scan1 = scan1->pip) {
1516 		scan2 = tdip;
1517 		while (scan2->pmp == tdip->pmp) {
1518 			if (scan1 == scan2) {
1519 				hammer2_inode_ref(scan1);
1520 				return(scan1);
1521 			}
1522 			scan2 = scan2->pip;
1523 			if (scan2 == NULL)
1524 				break;
1525 		}
1526 	}
1527 	panic("hammer2_inode_common_parent: no common parent %p %p\n",
1528 	      fdip, tdip);
1529 	/* NOT REACHED */
1530 	return(NULL);
1531 }
1532 
1533 /*
1534  * Synchronize the inode's frontend state with the chain state prior
1535  * to any explicit flush of the inode or any strategy write call.
1536  *
1537  * Called with a locked inode.
1538  */
1539 void
1540 hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
1541 		    hammer2_chain_t **chainp)
1542 {
1543 	hammer2_inode_data_t *ipdata;
1544 	hammer2_chain_t *parent;
1545 	hammer2_chain_t *chain;
1546 	hammer2_key_t lbase;
1547 
1548 	ipdata = &ip->chain->data->ipdata;
1549 
1550 	if (ip->flags & HAMMER2_INODE_MTIME) {
1551 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1552 		atomic_clear_int(&ip->flags, HAMMER2_INODE_MTIME);
1553 		ipdata->mtime = ip->mtime;
1554 	}
1555 	if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size < ipdata->size) {
1556 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1557 		ipdata->size = ip->size;
1558 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1559 
1560 		/*
1561 		 * We must delete any chains beyond the EOF.  The chain
1562 		 * straddling the EOF will be pending in the bioq.
1563 		 */
1564 		lbase = (ipdata->size + HAMMER2_PBUFMASK64) &
1565 			~HAMMER2_PBUFMASK64;
1566 		parent = hammer2_chain_lookup_init(ip->chain, 0);
1567 		chain = hammer2_chain_lookup(&parent,
1568 					     lbase, (hammer2_key_t)-1,
1569 					     HAMMER2_LOOKUP_NODATA);
1570 		while (chain) {
1571 			/*
1572 			 * Degenerate embedded case, nothing to loop on
1573 			 */
1574 			if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
1575 				hammer2_chain_unlock(chain);
1576 				break;
1577 			}
1578 			if (chain->bref.type == HAMMER2_BREF_TYPE_DATA) {
1579 				hammer2_chain_delete(trans, chain, 0);
1580 			}
1581 			chain = hammer2_chain_next(&parent, chain,
1582 						   lbase, (hammer2_key_t)-1,
1583 						   HAMMER2_LOOKUP_NODATA);
1584 		}
1585 		hammer2_chain_lookup_done(parent);
1586 	} else
1587 	if ((ip->flags & HAMMER2_INODE_RESIZED) && ip->size > ipdata->size) {
1588 		ipdata = hammer2_chain_modify_ip(trans, ip, chainp, 0);
1589 		ipdata->size = ip->size;
1590 		atomic_clear_int(&ip->flags, HAMMER2_INODE_RESIZED);
1591 
1592 		/*
1593 		 * When resizing larger we may not have any direct-data
1594 		 * available.
1595 		 */
1596 		if ((ipdata->op_flags & HAMMER2_OPFLAG_DIRECTDATA) &&
1597 		    ip->size > HAMMER2_EMBEDDED_BYTES) {
1598 			ipdata->op_flags &= ~HAMMER2_OPFLAG_DIRECTDATA;
1599 			bzero(&ipdata->u.blockset, sizeof(ipdata->u.blockset));
1600 		}
1601 	}
1602 }
1603